🔴 VectorDash - GPU Compute Bounties? LIVE Founder Interview w/Sharif Shameem

Don't concentrate on the finger or you will miss all that heavenly glory. Hello, hello.

Welcome, welcome. Oh my gosh, guys, I've got so many other things in the background that are just not correct. Oh, sorry. So as you know, last time we went live was for Unintended Crypto, so apologies for some of the mess you see in the background. We'll be doing some fixes there live to the, uh, to the backdrop of the show. But in the meantime, Sharif, without further ado, we've been sitting in the lobby for a couple minutes. Welcome to the show.

Hey, thanks. Hey, Seth. Thanks for having me.

Yeah, I'm glad we were able to finally make this happen. I know we were talking on Discord for a week or two, and I know that you've had your head underwater with this project. You have to have had your head underwater with this project. There's no way you could be making progress without that. So I can appreciate that it takes a while to get back to media requests. But this is, guys, this is kind of a different interview. We talked about the Bitcoin ETF yesterday, or the SEC's decision regarding the various Bitcoin ETFs yesterday. That was big news in the crypto world. So the Securities and Exchange Commission, just— this is breaking— just yesterday, they turned down some proposed changes to securities or exchange-traded funds on behalf of a couple of different entities. So New York Stock Exchange, they turned that down. And immediately we went live as a mining community and we were like, okay, what are we going to do? We checked the prices. Thankfully, crypto kind of held, so miners are— they're still going to keep mining on the GPUs, probably. But recently, over the last couple of months, profits have gone down. And the general sentiment with crypto miners is we kind of want to get these, you know, we get these little, these little workhorses earning us bounties some other way. And we had a different alt GPU compute bounty project on the show last week. They'll remain nameless, is your show. And, uh, and we had a couple of members of the community saying you should really bring VectorDash onto the show. So it's a big deal that you made some time for us. As you can see, the live viewer count is just going up because people are excited about this. So thank you for making time for this, means a lot.

Yeah, sure thing, no problem.

So let me, guys, let me turn off some of the, uh, the other like BitTube comments and other stuff because those aren't, those aren't going to be relevant. But let me start off by asking Sharif, tell us a little bit about yourself and then remind me the other team members' names. There's, uh, there's Abhishek and then there's a third, a third co-founder, right?

Right, yeah. So there's, uh, 3 of us. There's me, uh, Sharif, and there's Abhishek and Arbaz. Um, so I'm a student at University of Maryland. I'm studying computer Just finished up my junior year over there. Um, so one thing I noticed was that while I was working on like a bunch of side projects, like as a student, the compute for like working on machine learning side projects was really, really expensive. Like insanely expensive. I think it was like at the time $30 per day per GPU, per day of GPU compute on AWS, which is like the de facto standard for training machine learning models right now. And as a student, oftentimes, it can take a week to train or maybe even months if you're working on multiple projects. So you can see how that adds up really quickly. So I sort of realized that there are a lot of GPUs out there mining cryptocurrencies and they're only earning a couple dollars a day. So if you could figure out how to get those GPUs to actually train models for people working on these sort of AI side projects, You could, one, help out the people working on their projects, and two, give miners a much more profitable thing to do with their GPUs.

Yeah, no, for sure. And I mean, we definitely appreciate that. So today we're going to talk hopefully a little bit about the aims of VectorDash as a project and sort of the— if you've watched any of my shows or many of my interviews before, we're going to go through sort of a canonical set of questions that I would ask If I were to evaluate any tech startup of any kind, I would largely ask the same questions. Having been to a lot of pitch meetings for different tech startups and just various different business startups, these are the same kinds of questions I would ask any business. So if you feel picked on, don't feel special. But if you feel elevated or set on a pedestal, again, don't feel special. It's business as usual. We want to be able to evaluate your project, and if we're going to contribute GPU compute power, we want to know this is a good bet and this project has a future. So that's kind of the lens to look at this interview through for anybody who's just checking out this interview and for— well, and to help evaluate some of the answers to questions that come up. The community is here live as well, so they may have some specific questions about maybe fit and finish of the solution or maybe suggested requirements or required required compute power for folks who want to contribute, right?

Right.

That's a natural question. We'll see what else comes up, but typically, there are some semi-technical questions that come out of the audience. It's a semi-technical audience. You know who's watching largely? Clearly, you shared the link, but also from the Mind Your Biz community, there are a lot of small to mid-scale miners, and then also some mid-scale miners Who are consuming anywhere on the order of a half megawatt up to a couple of megawatts in their micro data centers that are largely GPU-based. So when they're looking at evaluating a solution, they're wondering, can this really scale? And is there a need? Genuinely, is there a need? So that's who's watching right now. And they're, in addition to wondering, okay, how far along is the project? Does it have legs? Does it have a future? They're also wondering, can I contribute meaningfully? Patrick Coombes says, handsome fella. He says you're a 9 out of 10, Sharif.

Thank you, Patrick.

Yeah, so Patrick's a good, good guy, good dude. Always great comments and always great, uh, always great color commentary on Twitter. So that kind of, kind of framing that, what I'm going to do is reveal, uh, reveal the VectorDash project site, and then if you want to help direct a little bit about what we can take a look at. Normally with a coin project, we try to evaluate a white paper, right? Because that's— it's Kind of de rigueur. It's what, what coin projects do. I don't know that VectorDash would, would put forth a white paper publicly, or if you can help guide us through it a little bit so that as questions come up, you can help direct me to find whatever publicly available answers there are so that the audience is now trained to go and seek out those answers on your website, that would be helpful as well.

Right, yeah. So, um, give you a quick run-through of the website. Um, so if you're looking to rent out your GPUs or host, um, we have a hosting section. It's vectordash.com/hosting.

Okay.

We also have a— sorry, I just messed up my video thing.

Oh, wow.

Okay, yeah, we also have— yeah, so if you make an account, you'll be able to log in and then you'll be able to view the available GPUs we have. At the moment, it's in use, but we are adding more in the next couple of days. We also have our pricing at vectordash.com/pricing. Our website's fairly simple, actually. I think we only have like 5 pages. Yeah, and if you want to get a quick overview of what VectorDash is as a host, I think the best page would be to— the best page to visit would be the hosting page.

Yeah.

It provides a quick overview of what you do, and at the bottom, there's actually a link to a Google Form where you can enter your machine's hardware and specs and what types of GPUs you have. And if we're interested in bringing it onto Backdraft Dash, we'll send you an email with more information.

I'm still waiting for my reply, by the way. Because I told you before we started the stream what I have here in the office, but it doesn't represent what I'm running elsewhere. So, yeah. Anyway, I kid. I know you guys have got a lot of submissions. Can you give us any kind of ballpark as to how many people have been like, please accept my GPUs?

Yeah. So I think this first started out when I posted a Reddit thread on r/GPUmining. I was expecting at most 4 comments and 2 of them being like, this was a stupid idea. But we've gotten a significant amount of GPUs or people with GPUs willing to rent them out. I can't give any specific figures on the exact amount we have, but it's definitely a decent amount.

That sounds like more than 4 people responded.

Definitely more than 4 people. Yeah. More than 5 even.

Oh my gosh. Okay. Well, you guys are done. Call— yeah, call the angel groups, call the angel networks, call the bank. You guys are set. That's what we call traction in my book. But that's great. Okay. So it's good to hear that the community response has largely been favorable. We've got questions coming up. Dragon, thanks so much for popping a question. Please put any question into the comment section, guys. I'll reserve your questions till a little later on. But please keep putting them into the comments section. I will go back and I'll pull them up later. So great questions already coming up. But I wanted to ask, as far as there being a problem that you saw that you wanted to solve, this is like the big startup accelerator question, right? And boiling it down to one sort of high-level statement, what was the big problem that you wanted to solve by creating VectorDash?

Right. So deep learning is expensive. That's about it. hard to make any significant project without being part of a large company like Google or Facebook or part of a large research organization like OpenAI. VectorDash is aiming to change that. Anyone across the world to be able to access the same resources that these large organizations have and put them towards training machine learning models, making significant AI advancements.

Got it. Okay. So you have a very specific buyer persona in mind, and that is the researcher or Or the organization that needs to consume, uh, they need to consume resources in order to do AI research then. And that's, that's who you've narrowly focused this solution on.

That's correct, yeah.

Okay, awesome. Cool. Just want to make sure that I understood. Yeah, I, I'll occasionally ask clarifying questions, and, and mostly it's just because I'm, I'm underslept like all the time. But, but also, you know, yeah, yeah, pretty much you could say it 3 different ways, then You know the audience understands. So, why then start with this project as opposed to— I don't know. Why start with this project as opposed to maybe trying to find a different open source community or a different project that already had momentum or team or resources and then becoming a vocal part of that community or vocal part of that project?

Yes. That's actually what I was initially planning on doing. My initial goal wasn't to start Backdash at the time. When I was working on my projects, I sort of thought something like this would already exist. I spent like a full month and a half Googling, like looking for viable projects I could use. But there were some out there, but none really actually worked. Like, right off the top of my head, I think Gollum was the one everyone pointed me to at the time. But there's really no way to train a machine learning model on Gollum right now. I realized, huh, I could probably just build this out myself and That's what I decided to do.

Oh, awesome. Okay. So in just a moment, we'll talk about team and background, because I think that's super important. And I think that a lot of different— a lot of new crypto investors and a lot of miners that are getting into the space and becoming sort of like de facto investors, because you're ponying up a lot of CapEx in order to create these micro data centers that are GPU-driven. And so Sure, you're not surrendering that money to a project, but you are surrendering those resources, so it is a form of investment. I think a lot of crypto miners and crypto speculators, they neglect to ask about team and figure out, why is this person qualified? Where are they going? We'll talk about that in a minute, but can I ask, as far as solving this problem, when you first got underway, just in terms of the very first hurdle? What was the biggest hurdle from a technical standpoint, and then maybe even from a governance standpoint, and forming a brand new organization or a new company, a startup? What was the biggest hurdle that you had to overcome to get VectorDash?

Right, yes. Yeah, so I think the biggest problem when I first started it was, was, was actually like And people was thinking about whether people wanted to use something like this or not, because AWS was huge. Like, they still are huge. They're massive. And there wasn't another project doing this. So I sort of thought, huh, no one else must be doing this for a good reason. And I sort of like, I was, I was filled with doubt as to whether or not people actually— on both the AI researcher side and both on the crypto miner side as well. So I guess sort of getting that initial interest from both sides put my doubts to rest, which was pretty great.

Awesome. Okay, so it sounds like then you went the path— and I've got to ask, do you have a background in startups? Have you already worked with any other startups before?

I've not worked for any other startups before, but I have been programming for quite a while. I wrote my first program when I was like 11 or 12.

So like 2 years ago. Got it. I'm messing with you. I'm messing with you. Sorry. I know you get a lot of it. I know. It's part of the interview process. It's part of PR, so sue me. No, that's awesome, though. It sounds like then, yes, you're an engineer at heart. You want to solve problems.

Yeah, yeah.

I promise I'm not an ageist, by the way. That was super messed up. I'm totally not into that. I just like, I haven't done that in an interview yet, and it seemed like, what the heck? Again, forgive me, audience, especially VectorDash supporters. Sharif is awesome. And clearly knows his stuff, otherwise he wouldn't have taken on this challenge. So that's awesome. So no previous startup experience per se, but clearly— here's where I was going with that comment before. It's a roundabout way of giving you a compliment. Very few startups, certainly in the crypto space, do proper market validation. And that's a big startup problem in general, is doing market validation and knowing for sure, am I peddling a solution to a problem people really have?

Yeah.

Right? As opposed to like, well, this seems neat to me, so why don't I just do it? So what did you find as you got into that process of discovery and you started looking for validation of this concept? Like, where did you go? Who did you ask? And what were their responses like?

Right. Yeah. So VectorDash at its core is sort of like a marketplace. We have the providers, so people who own the GPUs, and we have the people who consume the GPU resources. And these people are sort of machine learning researchers or AI developers, anyone working with compute-heavy machine learning. So I was sort of more familiar with the machine learning side of things, actually. I wasn't originally a crypto miner. I guess I'm still not. I was more familiar with what the requirements were for training these machine learning models on these GPUs. So sort of working with frameworks like TensorFlow and PyTorch. accelerated libraries. So I would just talk to people who I knew were working on similar projects. So people who were like, like working on computer vision stuff. You know, I had a friend who was working on some recurrent neural network research. And I asked him about VectorDash, or like what it could be. And he was like, yeah, sure, I'd use it. I'm paying like a crap ton of money to Google already. So why not just pay less for the same thing? And it wasn't— I was more worried about the GPU miner side, I guess. I wasn't sure if they'd be willing to One, um, like install Ubuntu, which is like a requirement for our, uh, for our desktop client because we can't run it on Windows.

Oh yeah, here's the thing though, my audience is tired of hearing me say it, but diehard miners have been running Linux the entire time anyway.

Yeah, I just— I didn't know that at the time. I thought everyone was running Windows and like NiceHash, and I was like, oh God, this won't work on Windows.

Awesome. Okay, so then it sounds like then What did you have to do in order to assuage the early concerns of your co-founders and of anybody involved in the project early on that miners could be involved? Or what did you discover when you were doing that market validation that helped you be comfortable moving forward?

Yeah. So I guess what we did was we just put up a Google Form on our website and we just let people enter out what hardware they'd be willing to list on VectorDash. Validated our hypothesis that people are willing to list their machines to train these machine learning models. They are willing to, one, take a risk on this unproven project, and two, they are willing to do something besides mining cryptocurrencies.

Okay. Awesome. So the numbers speak for themselves, it sounds like. And have you been able to validate the actual microdata center build-out that is being alleged by miners who fill out the interest form?

Yeah, so we've actually had a lot of people willing to build machines specifically for VectorDash itself rather than like for mining. So some people have been willing to use their mining rigs to become sort of VectorDash hosting rigs and to build rigs from scratch just to host on VectorDash and not necessarily mine at all.

Okay, yeah, awesome. So I'm among those proponents of building a slightly higher grade of mining machine. So it's not like you're not looking at data center quality builds with redundant power supplies and Xeon or server-class CPUs, but just a cut above what's traditionally, for the last few years, been used for mining. It's funny, I catch myself saying stuff like that with regards to mining. It's like the only tradition here is that we threw tradition out the window and started choking GPUs through single-lane PCIe connections. Um, so it's, I don't know, kind of weird, but awesome. It sounds like then you've heard good feedback. And where I was going with the question regarding, uh, regarding validating some of the claims of miners is that miners are a bit like fishermen. In the private forums, they're always like, I caught a, you know, I caught a bounty this big, or like, I, I discovered this many blocks with this rig, or I used this this particular mining software and I got this hash rate. And every single time they tell the story, the number gets bigger and bigger. Right.

Right?

So that's why I'm like, I'm always a little dubious when people are like, yeah, my mining rig is just amazing. I got this hash rate. It's like, prove it.

Yeah, yeah.

And it can be— proof can be harder to come by, in my opinion. But that's awesome. It's awesome that it's been a really positive community response. I know in our community we've had a lot of people who are excited by the announcement of VectorDash because they've heard me kind of ramble on about like, no guys, build up your rigs a little bit more, be prepared for alternate compute bounties because they should be coming any day now. And some other projects that are interested in it as well. So that's awesome. Glad to hear that, that Market validation has gone well so far and the response has been good. What new problems are you finding that you want to solve now? At first, you mentioned market validation. And as far as the actual project itself and distributing the workload and the technology, what are some of the problems that you have on the roadmap and that you've been working on recently?

Yeah. So what we want to do is sort of link up people who request these GPUs, so these researchers, with machines geographically located near them. So we have locations all across the world, so it wouldn't really make sense for someone located in Washington, D.C. to use a machine in like Beijing or Shanghai. So what we're doing right now is sort of calculating what machines are closest to them based on ping and then assigning them those machines. So you can be guaranteed that the machine you're getting is up to par. It is That it is what you selected, and it's also relatively close to you as well. So in the same way that when people spin up a server on AWS or Google Cloud, they usually select the data center closest to them. But what we've noticed with Vectordash is that sometimes the GPU you're using can be in your neighborhood. It can be like in the city over. It's not like 100 or 1,000 miles away. It's usually less than 50 miles.

Nice. Oh, that's amazing. Yeah, that's great. So, um, all right, so it sounds like then recently then you've been focused on the, the problem of linking up buyers with sellers, with hosts, right?

Yeah, so we have it, we have a preliminary system set up right now where we don't look at geographic location at all. We just sort of ignore it. Um, so right now we're focused on building out this sort of, uh, this new feature of machines, um, and taking that into account when linking up, uh, Okay.

Got it. And we'll get back around to some other technical questions. And again, the audience is semi-technical. Not many of us are programmers. Some of us understand the frameworks well enough, and some of us understand network and security well enough to be able to kind of grok some of the answers that you might have already come up with regarding some of the technology. But we're not programmers ourselves. So bear with us. That's the big caveat there. But as far as the team, let me jump back to the about page. And let me ask you to give us— oh, wow.

Hmm.

It's funny. Navigation. The about page. That's funny. The about tab. I guess— no, it's never been on top navigation, huh? It's always been at the bottom.

I think it's in the footer. Okay.

Yeah. Yeah. Thank you. But as far as, um, as far as having an about or a team page, can you tell us a little bit about, uh, Arbaz and Abhishek and kind of what they're, what they're studying or what their background is? Are they, uh, are they students as well or are they, are they already moved on?

Yes.

To, uh, to other positions?

Um, Arbaz was one of my close friends. Uh, he still is. I met him like first grade, so I've known this dude for like a long time. Um, yeah, he, he goes to UMD with me. went because he graduated this semester. He's also studying computer science. And Abhishek is another friend of mine. He goes to Carnegie Mellon University. Both of us are still students there. I go to University of Maryland and Abhishek is a student there at Carnegie Mellon. But Abhishek is— I'm sorry, Arbaz has graduated right now. So 2 of us are students and one of us just graduated.

Okay, awesome. And it sounds like then Abhishek's got right there that he's studying computer science and computational finance, which is amazing that it's a thing now, but that's a good thing that it's That it's a proper course of study. Otherwise, universities would be quickly losing relevance.

Right, right.

Yeah. So it's awesome. So as far as core skill sets, what do you each bring to the team? And is there sort of a rough division of labor that you have right now? And if so, what does that look like?

Yeah. So for the most part, Abhishek is responsible for bringing hosts on board and handling machines or any queries they have. Menarbaz and I are more focused on sort of the the machine learning and the research side of the market. So it's sort of split in this 2-1 sort of way, where we'll go out and try and find people to use the GPUs and Abhishek will sort of bring them on board.

Got it. Okay. Yeah, very, very cool. So you guys are doing quite a bit of sales. Kudos to you, by the way, also for doing that. A lot of new startups, they neglect to assign themselves the role of salesperson. It's pretty important, right? That whole like if you don't, if you're not cash flow positive, you got problems down, you know, looming. So like I know that sounds like a weird thing to to call out, but I've consulted too many startups that are just waiting for like an angel to step in, you know. So so that's awesome to see. So it sounds like you guys have you got a bit of a division of labor going on there. As far as as far as what brought you all onto this project, what spurred the interest of each individual team member and Kind of like, what, what's— what goes on in the mind of a VectorDash contributor/dev? And what gets you up out of bed each morning saying like, all right, because it's startup time, right? We're running lean just on our burn rate and contributing with what little time we have. Like, what gets you up out of bed thinking like, thinking like, yes, VectorDash today, I'm so excited?

Yeah, so I guess it's— I'm gonna nerd out here for a bit, but I think artificial intelligence is really, really cool. And we're sort of at this inflection point right now. We've been progressing pretty good these past couple decades, but these past few years, we've made significant progress both in machine learning and deep learning. And there's that original chart. I think my sister's making toast downstairs or something. Sorry about that noise.

Yeah, I was going to say, if you were anywhere near a mining rig, I would say you're probably using a SATA cable to power your router.

I know she's probably just burning her toast. It's fine. But yeah, so we sort of realized that the possibilities, the things you could do with artificial intelligence, they're absolutely insane. You can do so much things. You can treat cancer. You can classify X-rays. You can build these sort of chatbots that are just like humans, which is one of my original projects.

So natural language processing and all the rest.

Yeah, right. these things were accessible to you unless you were part of like a larger organization like Google or Facebook or any other company which had like these mass amounts of compute resources and like these mass teams of researchers. I have this saying and it's sort of like, what was it? It's on the front page of our website if you go to it.

Sure.

Near the bottom. It's sort of that like talent is evenly distributed across the world, but then you sort of have resources not being evenly So you could have someone in Kenya who's just as smart as that Google engineer, but the thing is he doesn't have access to the same resources that Google engineer does. What VectorDash aims to do is to give, give these people across the world the same resources a Facebook or Google engineer might have at pennies. We want them to be able to advance the field of artificial intelligence without having to worry about spending a significant amount of money on compute resources.

Rackley says that he's very interested in a very specific field of AI development for interpersonal relationships. I don't know. I don't know, Rackley. Your princess is in another castle. You have to wait until they've gone mainnet and we're all helping them with more projects. But awesome. Okay. So it sounds like then the thing that is helping in terms of being sort of a cultural glue and sort of a Sort of a vision for the team is to help distribute resources.

Correct. Yeah. So we want— yeah, we want people to work on harder problems. And people aren't going to be working on harder problems unless they have access to more resources to work on those problems. So at the end of the day, the main goal with VectorDash is to advance AI. That's it. We— and we realized the best way to do this is to distribute resources even when you're across the world and make really powerful GPUs. Like, people will have 8 GPU rigs. That's actually a mini supercomputer right there. And if you let someone who knows, um, who knows what they're doing with a, with a machine learning model, you give them access to that rig, they can do insane things. Um, like, it's absolutely crazy. And the fact that it's just sitting in a basement, like, crunching away hashes, it just, just kind of irks me a little bit, don't you think?

Um, well, if you, if you tuned in any of my other coin projects, you, you've seen me argue with half-billion-dollar market cap coins about just that point. So I agree with you. I do. Yeah. And these are people who are like, uh, you know, they— it's funny how crypto kind of came out of a very anarcho-capitalist libertarian set of values and sort of like pulling down power from the centralized powers and, you know, sort of, uh, gosh, what is it, um, sort of Promethean act of like stealing fire from the gods. And, um, But now there are these coin projects that are like, oh no, no, our market cap is too big. We can't, we can't just let you guys have control. It's like, wow, it's, uh, it's all just a cycle, isn't it? Um, but this is— that's super cool. That is super cool. And I agree with you. Yeah, there's way too many, way too many, um, cycles that have been scavenged, scavenged for, for hashing when the hardware is far better suited for other stuff.

Right, right.

Yeah. And it turns out that with GPUs also, in terms of the cryptographic hashing algorithms that are used, a lot of them don't even properly utilize the hardware. They don't even properly utilize each side of the GPU. And certainly, a lot of them aren't being written with the CUDA libraries to make sure that the algorithms are being properly optimized. Some of them are. Some of the miners are doing better at that, but even still, full utilization of the GPU with mining. It's not there. So there's not really good efficiency there. And then with application, like you're saying, how sad that a portion of our hash power, the 22 million and counting GPUs in the world, that it isn't going towards cancer research or building Rackley a better girlfriend. Sorry, Rackley. We can't make Rome wasn't built in a day. These guys are working on it. So yeah, awesome. Totally agreed. And a lot of the community from, from Mine Your Biz, ironically, a lot of us talk about mining, but a lot of us are agreed as well. We want to be able to deploy our GPUs in a different environment and earn different bounties also, because for many of us it is a business decision. We, we do have— we keep books that are dedicated to just our mining operations, or, you know, we track our expenses and income and And, uh, bare minimum, so the IRS doesn't nail us at the end of the year, but also because some of us have really scaled up and we want to be able to contribute to projects that are, you know, that are gainful. And if they can be meaningful also, that's cool. It's definitely cool. But for a lot of miners, it's kind of secondary. It has to be profitable first, right?

Yeah. So that was one thing, um, when I was first starting Vetradash, one of the core requirements for me to actually start working on it was that it would have to be better for both sides of the market. So for people who are actually to train these machine learning models and work on these projects, it would have to be significantly cheaper than Amazon or Google. The threshold I set was 10x. It had to be 10x cheaper or I wasn't going to do it. And on the other side of the marketplace, the GPU miners, they had to be earning significantly more than what they would make mining the most profitable cryptocurrency. So I remember like every day for a month, I would check whattomine.com and I would like sort it by like the most profitable cryptocurrency. And I would like, I would look at the price and I would like compare it to Amazon's GPU price or the GPU costs for a day. There's this significant gap there. If you can figure out how to close that gap and give those resources to people who actually want to use them for meaningful projects, one, you'd be helping out both sides of the market, and two, you'd be advancing AI a non-trivial amount.

Awesome. I love how you say it. We don't want to name numbers, but non-trivial for sure. You can tell you're talking to an engineer when that term comes up. Great. So hey, Richard, thank you for saying hello in the comments. We've got a couple of mine operators in the comments who I think would qualify in terms of who you're looking for and in terms of the build-out of their mines, who are just saying hello in the comments. But we've got a few comments or questions that are coming from the audience now. Normally, I would ask any coin project that I interview, I would ask them a series of questions regarding hardware. There's a There's a big movement within the hardware manufacturing scene to move towards application-specific integrated circuits. And so Bitmain in particular, really big player in that space. I'm sure you've heard of them. I'm sure you've read about them. And I'm sure if you're into mining that you've even considered maybe a hardware purchase with Bitmain. But there's a bit of a backlash within the mining community, GPU crypto mining community, because of the sunk costs. that many of us have into GPUs. And because we know that because they're GPUs and because there are great libraries for OpenCL, CUDA, uh, what is it, AMD with ROCm now, um, and, and then all these other, these other projects that can be adapted to the GPU hardware, that they're just, they're a lot more flexible than the ASICs could ever be. So we feel like that was a good choice. But when profits dip, it's like, oh, a lot of miners lose faith. So, and so they, um, with recent advancements in ASICs and the ability of the ASIC manufacturers to create hardware that can do memory-hard algorithms that were previously thought ASIC-resistant, it was commonly believed that only GPUs or CPUs could do these hashing algorithms. That was a bit of a shakeup in the mining scene. And, um, So with coin projects, I've been kind of holding their feet to the fire, like, hey, if you started with the objective of, uh, of keeping your coins GPU mineable, what are you doing to assure that that's still the case? And we've had kind of a— basically it's been kind of a beatdown session where it's like, hey, we've got GPUs, how are you going to help us keep them profitable? In your case, that's like the whole goal, right? So a whole different line of questions here. Uh, but the questions I'm getting from the audience have to do, uh, have to do with— and we've got a ton that are, that are loaded up now in the comment section— but they have to do with requirements for participation. What would be the minimum required spec for a system or, or for a, a mini mine to flip into utilizing VectorDash, or you utilizing them?

Right, yes. So, um, specs are pretty simple. We do have slightly higher specs required than a typical mining rig would have Just because deep learning is intensive. It's keeping the GPU saturated most of the time compared to hashing. So right off the top of my head, we do have CPU requirements: one physical CPU core per GPU on your machine. So if you have a dual-core CPU, you can't host— you can't host 4 GPUs on it, but you'll be able to host 2 of those GPUs. We also have PCIe lane requirements. So 16x or 8x PCIe lanes. And this is also due to the fact that deep learning is really, really bandwidth intensive. You're feeding data to the GPU constantly. And if you have a 1x lane, you'll just significantly hinder performance.

Is there a revision of the PCIe spec that you require as well? PCIe 3.0?

Yeah, yeah, 3.0. And then RAM requirements for lower-end GPUs like the 1060, We have a RAM requirement of around 4 gigabytes, and for everything above the 1060, we have a RAM requirement of 8 gigabytes. So, just to be clear, this is like the RAM of your computer, not the GPU's VRAM. So, if you have a machine with like 16 gigs of RAM, you'll be able to host 2 of the GPUs. What else is there? Oh yeah, disk space. So, disk space, this is for people who load the dataset onto the GPU. They need somewhere to load it on their computer. So just—

and I apologize, clarifying question— when you say 1 CPU thread per GPU, do you mean a physical core, or can that be a virtual core like Hyper-Threading?

Yeah, we mean 1 core. So if you have like a quad-core Intel i5— I think the latest generation is 6 cores, actually— so if you have like a 6-core Intel i5 this generation, you'll be able to host 6 different GPUs on that machine.

Okay, so what about a Core i7 where hyperthreading is enabled? These need to be these need to be physical cores, not logical.

Right, they need to be physical cores, not logical or hyperthreaded cores.

Got it. Okay.

Yeah, and um disk space requirement is 50 gigs per GPU.

Got it. Okay, so we're just taking note of that. BDF lulls at BDF saying yeah PCIe 3.0. With 8 lanes. RIP. RIP my systems. Yeah, sorry, BDF. Sorry. This is not your grandfather's hashing machine. So, okay. Yes, Rackly, we know that PCIe 5 is in the works. Call me when it's commercially available. Awesome. Okay. So that helps a lot. That helps a whole lot to be able to understand just how bandwidth intensive it is. And it sounds like memory again, just to recap what you said, 8 gigs of system RAM per GPU. That, uh, so you need a full physical core and 8 gigs paired to that physical core per GPU. And the GPUs themselves, is there a, uh, is there a minimum spec for the GPU? So you mentioned the GTX 1060. Would the 3 gigabyte model be okay? Would the 6 gigabyte be preferred?

Right. So we, we let people actually pick which GPUs they prefer. We did some benchmarks on the 1060 3GB, and when training machine learning models, you get quite a bit of out-of-memory errors. So you have the dataset and it's not, it's not being able to fit in the GPU, or at least the batch. So the way it works is that you load the dataset in batches into the GPU, and if the batch doesn't fit in, you have to either decrease the batch size or get a larger GPU. So we sort of realized that just Setting the lower boundary on GPUs as the 1060 6 GB was an option, just because it seems fair. Most machine learning researchers probably wouldn't use the 1060 3 GB, just because the GPU's VRAM is really important.

Got it. Okay. And is there— so is there a minimum for the GPU's VRAM?

GPU's VRAM? So most GPU models have this VRAM across the model. So like all 1080 Tis will have like 11 GB, I think, and like Old Titan XPs will have like 12 gigabytes. But like the lower boundary of GPUs is NVIDIA 10 series GPUs. So like the 1060, 1060 Ti, 1070, 1070 Ti, all the way up to the 1080 Ti. And then Titan GPUs. So like the Titan V, Titan X.

Got it. Okay. So, all right. Realistically then, you want us to totally spec out then, if possible, I—

yeah, um, well, for the most part, if you do the math, it's actually not that much more expensive. Um, like, most people already have like a dual-core or quad-core CPU that can host it, um, unless they're running like a Celeron or a Pentium, and they're out of luck. Um, but for the most part, um, I guess the largest— the most expensive thing would either be the motherboard for the PCIe lanes or the RAM, because RAM is really, really expensive now for some reason. Yeah.

Yes, yes it is.

Yeah, so while we do have like a bit of requirements than like the traditional mining rig would have, um, I've also made sure we pay out significantly more than mining does as well. So if you go to the hosting page, actually, I think there's a list of our payouts.

Right, and I've got a question from the comments section here, and I appreciate it being asked. XBitCo asks, is this data stated somewhere on the website? The minimum spec requirement?

So when we look at the spreadsheet and if we come across machines that match the requirements, we send them an email with more details. So just because of the large number of responses we've gotten, we're still testing out different specs, different sort of hardware combinations, seeing what works best. So these specs are— there are— we're still working on them. figuring out what exactly the best specs are for both machine learning researchers and for GPU miners. You sort of want to strike that even balance between both of them where you don't have these GPU miners building out data center servers and you also don't have these machine learning researchers training on 1x PCIe lanes. You got to hit the right balance between the two.

Got it. That makes sense. So it sounds like you want to be as inclusive as possible then for the mining community and grow your inventory of hosts for new clients. It's brilliant. I mean, you can't argue with that. Makes a lot of sense. Let me jump back into the comments to look for some other questions. I've got one from Robin who's asking— Come on, Robin, be nice. Sharif is a guest. Let me filter this one just a little bit. But Robin essentially asks why, if there's a data center operator that has a significant number of GPUs and they know they can achieve the bandwidth necessary, why would they work with you As opposed— as something of a middleman, as opposed to going direct to clients.

Right.

What's the value proposition of working with VectorDash?

Yeah, so what we do is not only do we offer like this sort of platform where we're like, we'll handle security, we'll handle infrastructure, we'll also handle the sourcing of clients. So we're actually working with like these large companies. We're actually working with the University of Maryland to get them to use VectorDash GPUs for their machine learning courses. and other universities across the East Coast as well. So not only are we out here providing you the infrastructure, we're also planning on sourcing customers and pointing them to your hardware. So I guess you could say, like, actually building out the platform is only like a third of the problem. The other, like, two-thirds are actually getting the machine learning researchers to switch away from AWS to this sort of, like, untrusted cloud provider.

Got it. No, that makes sense. That really does. And as far as working within education, I apologize, there's a little bit of lag in the video stream. You mentioned you're working with the University of Maryland, and do you have any soft commitments from any other universities also to work with you?

Some are in the speaks. I can't really share at the moment, but we are working with other universities.

Okay, got it. So there's definitely an outreach effort underway on your part.

Right, totally, yeah.

Okay, awesome. And if you had to, at current usage levels, issue a ballpark statement about the utilization or the capacity that the University of Maryland needs, it seems like you're most familiar with their program.

Mm-hmm.

What is their consumption like today? How many GPUs could they potentially use If they were to go all vector today, right?

So, um, the figure I got was— so that you— I mean, UMD uses AWS for their GPUs, um, so students will get GPU credits to use for their machine learning courses.

Which, by the way, are very old. AWS, Google, Azure, IBM, all their GPUs are old.

Yeah, so yeah, they have like— um, they've actually some more recent ones, but for the most part, most people use K80s. Um, so they have $50,000 per course, and there are multiple machine learning courses, and this is per semester. So you do the math and translate how much does $50,000 in AWS translate over to Vetradash. What we wanted to do—

Well, you said you wanted to be an order of magnitude less, so you're offering to pay us $5,000.

Well, if you think about it, it doesn't really translate. They don't have a set requirement, right? they have a budget, you could say. So it wouldn't be that they're paying 10x less, but rather they're getting 10x more compute for the same price, if that makes any sense.

Well, it— yes, yes, it does. It does. So in practical terms, we'll have to, as a community, we'll have to figure out what that actually means for us in terms of what we're providing. Um, because very few of us— like, it became unprofitable to mine on, uh, on VPS A while ago. Yeah, it became very unprofitable to mine on VPS a long while ago unless you can provision some of the latest stuff. And yeah, I mean, you're right, there are a couple of cloud providers that have upgraded some of their GPU. But yeah, so I get it. The mining economics are attractive from that standpoint, but the bandwidth I think is going to be a problem that comes— that you hit with a lot of the inventory that's out there right now.

Right. Yeah. We're also working with them. We're like running benchmarks on multiple different configurations. So we're trying to strike just the right balance such that most miners can still use their mining rigs, if not all of them for some of the higher-end rigs. Even if it's stuff for like lower, lower-end machine learning models. People are training like really intensive machine learning models right now. But in the future, people might not have that same bandwidth requirement. So we're really trying to figure out what's best for both. Um, for both parties here.

Okay, awesome. Got a couple of questions from the, the comment section about, uh, current onboarding efforts. And as far as end users, anybody who's been onboarded into early beta, are there any— or rather, are there any hosts that have been onboarded to date as far as doing tests with you? Or is it primarily tests that you're doing?

Right, yeah. So we have onboarded a significant amount of hosts. What we're— who we're working with right now are primarily users with like large numbers of GPUs rather than the individual miners. Just because, one, it's much easier to test when you have one point of contact versus 100 or 1,000. And two, the turnaround time is much quicker in terms of development. So if we have a problem, we'll be like, hey, could you run the newer version of our client? And they'll be sure, and they'll load it across 100 different machines. Whereas with developing with 1,000 different people, it's a much slower turnaround time. So it's just easier to develop when you have fewer people with lots of rigs rather than lots of people with fewer rigs. Sure.

No, it makes sense. I mean, as far as it being a procedural workflow, I mean, smaller data center operators or mid-scale miners, they have processes in place to change out mining images fast, and they've got better network fabric, and they understand how their mine operates better than somebody that's running 3 rigs in the basement. So that makes sense. Definitely does make sense. As far as there being a As far as there being an opportunity for someone to be involved in early beta efforts now, what would— what kind of hardware would they have to be able to provide at this phase to be considered to work early on? Right. Yeah. So as a host—

so what we do is every few days we just go through the form and look at the submissions. And we— it's really hard to develop a sort of software that works across multiple different machines, multiple different GPUs. CPU motherboard combinations, and that works seamlessly across all those. So what we do right now is we have different tests and different development efforts we're running on different sort of like machine CPU motherboard combinations. So what we need to figure out, hey, will this, will this new version work on AMD Threadripper CPU with like this motherboard? We'll look for a host who has that matching requirement, and then we'll ask them, we'll reach them out, hey, Would you be willing to test this newer version of our clients? And yeah, that's pretty much how it goes for the most part. We are looking to get more AMD CPUs. We have a lot of Intel CPUs, but we want to test more on AMD CPUs right now.

Got it. Okay, so there's— guys, there's a participation signal you got direct from Sharif. Let me give you an ear horn. Wait, come on now. My soundboard's been acting up.

Good grief.

Okay, there you go. Air horn for Sharif. Guys, there's a, there's a, a contributor signal right there, or a hosting signal right there. They're looking for AMD, but specifically you're looking for Threadripper, high core count, right? AMD. Yeah, awesome. That makes sense. Yeah, for a host device, definitely makes sense based on the specs that you listed before. Somebody is asking about, is location of GPU important or not? Can, can people be reasonably considered for the early beta if they are, for example, in Eastern Europe right now, or would that be difficult today for you to work with?

Um, so location actually, it, it matters if there are people who are willing to use your machines in that location, right? So if there is a mining rig in Alaska or Antarctica per se, I'm willing to bet there aren't going to be that many machine learning researchers in those locations. Are you sure?

A lot of oil and gas people out in those areas.

That's true. That's true. Who knows? I could be wrong. So location, when we're developing for the private beta, we're really looking at location, but definitely once we've launched the public version, location will be very important because if you're the only GPU operator in a country like I don't know. There are a lot of machine learning researchers in China, but there are also lots of miners there as well. So you could be the sole provider for GPU resources for all the machine learning researchers in here if you're the only one providing resources there. So geographic location is important to the extent that there are people willing to use the hardware there.

Awesome. Okay, great. So it depends on demand then.

Correct.

Yeah. So you'll be matched to whatever the demand heat map is. Very cool. Somebody asks also, is this solution bottlenecked by connection speeds? And I don't think they're talking about PCIe lanes or PCIe speed. I think they might be talking about network speed. Is the dataset just loaded onto the host system and then work done there and then results sent back? Or what does that look like by way of network connection?

Yeah. So the average workflow is sort of, it's It's pretty simple. The machine learning researcher loads 2 things onto the GPU machine. They load one, the actual model. So this is like an untrained model, an untrained neural network. And 2, they load the dataset. This dataset can range anywhere between a couple of megabytes to upwards of like 10 gigabytes. There are massive datasets out there, like upwards of like 20 terabytes. But for the most part, those people are like, they're probably already using AWS and Google. And they have the network capacity to support these like massive datasets. Like ImageNet is one of them. And it's, I think it's like over a million different images. It's crazy. Um, so internet speeds, I guess you could say they do matter as long as it's not awful. So if you have like 1 megabit per second up and like 2 megabits per second down, then most people probably aren't going to use your rig. But for the most part, Most broadband connections will be fine. I think like 25 megabits per second up and down would be all right. We don't need like super high, like 2 gigabit internet connections per GPU rig. That's not feasible, and most people don't even need it.

Got it. Okay. That makes sense. And as far as your own network configuration or your own hardware configuration, are you looking for hosts that have already— have a preconfigured cluster? Or would that be helpful? Or are you looking for individual host devices to be like one motherboard, one CPU, the RAM specified before, some GPUs?

So do you mean in reference to like the private beta or—

In reference to what's moving forward with that as a host contributor.

Right, right. Yeah. Right now we definitely want to utilize all machines. We don't want machines to be underutilized. So right now in the early stages, we're primarily working with like these larger farms, but we do plan on getting all the machines online. And what we've noticed is that some people are actually using these GPUs for non-machine learning purposes. I've actually had someone reach out to me a few days ago and they're like, hey, can I use these GPUs for cloud gaming? Because there's a GPU located in the next city over. And I versus Amazon or NVIDIA data centers 2,000 miles away. Can I use that GPU for cloud gaming? I mean, I hadn't thought of that, but that's actually a really, really interesting question. Because if you think about it, this wide distribution of GPUs all across the world, and while AI is pretty great, games are awesome too. I'm a huge gamer myself. How awesome would it be if you could host your own cloud gaming service on your GPUs? That'd be great. So we do plan on letting We do plan on utilizing all the GPUs. That's the—

Got it. So I'll rephrase my question to be a little more specific. If a mining— like if a data center operator has already preconfigured their own multi-node cluster, would that be helpful?

Oh, so you mean an interconnected cluster of machines?

Correct. I'm sorry, I just took for granted that it would mean the same thing to you.

No, well, for the most part, um, inter-GPUs, multiple GPU motherboards are helpful, but connections between different machines, we don't use connections between different machines at the moment. So we take— we assume each machine is an individual entity and we don't assume it has any other connections with any other machines in the same location. So for the most part, probably not, but moving forward, um, Who knows? It could be useful.

For now, it's probably not. Okay. I'd like to propose— without introducing too much scope creep, because it sounds like you're already getting distracted by the shiny game hosting, which, I mean, I'm not going to lie, it's pretty cool. So I get it. So without shiny object syndrome creeping in, I'd like to propose that you perhaps allow for hosts to run their own multi-node cluster. That might help to overcome some limitations that we have with our current hardware builds, because we can then have one of those nodes meet some of the requirements in terms of coordinating jobs with thread count and with RAM. Just throwing it out there.

Awesome. Yeah, look into it more. It seems really interesting.

Okay, cool. That, and I mean, you know, I speak for the trees, right? These are— I advocate for the mining groups that are that are kind of saddled with some of the systems that they have, so, and they'd love to participate in something that's maybe more lucrative without like, without exploding their CapEx budget. Right.

Right.

So, so yeah, maybe adding one node to, you know, an otherwise decent 4 or 5 node setup, that might, if that's possible, I know that would, that would help you onboard more of my community for sure. So another question that came up in the comments section Wood. I'm trying to filter myself through some of the trolling because it's just fun. It's in good nature. So Patrick is asking, have you thought about lowering the minimum to 20 GPUs so we can test the profitability rather than going all in on 100?

So contrary to popular belief, we don't have a minimum GPU limit. You can list as many as you want to. I do remember sending out an email a while ago saying if you had 100 GPUs, contact us directly. That was probably not worded in the best way. It was meant to be like, hey, a lot of GPUs, just let us know and we'll see what we can do. But it's definitely not a requirement. We have hosts right now with like 2 machines online with 4 GPUs each. And so it's easier in that we're working with larger hosts to develop, but it's definitely not a requirement. Yeah, it's— I should have worded that email much better at the time. That was my mistake.

Oh, hey. All right. So you guys heard it. No minimums. Just make sure that you're up to spec. Again, making— like, not sure if trolling or real question. Give me just a moment here. And you've answered a lot of it, actually, by talking about specs.

Hmm.

All right, so I've got— I'll just— I'll toss you this hot potato. Somebody says it's like Golem or Sonem Project. Hosts would make a lot, but there's not enough demand for them to make a profit.

Right, yeah.

So I don't know if I agree with that. What do you say?

This is actually a really great question. I really like this question. When I first started Vectordash, I was looking at Gollum and Sonnen. I really wanted to use something like Vectordash at the time, but it did not exist. I was trying to train a model on Gollum, but it's, it's literally not possible. Like, you can't do it. Um, I was willing to like—

Hey, why is that?

Um, right now I think you can only render Blender files on Gollum. Um, they have mentioned that they do plan on like adding more features in the future. Like right now, you can only render Blender files. And I've never used Blender before, so I have no interest.

Well, Cycles Renderer is free and open source. So I think they're just— they're saving themselves the hard work of trying to knock on Autodesk's door and being told no repeatedly for the next 6 months. I've worked with startups that have had that problem. So it's a real struggle to get software partners on board.

Right. Yeah. So one thing I've noticed is that most of these projects, or almost all of them, are focused primarily on the GPU side of the market, and they pay no attention to the people who are actually willing to pay. So coming from the, coming from the side of the market, who, like, coming from as someone who wants to train machine learning models on GPU, I'm like, most of these, like, distributed systems just, like, outright don't work. Because I've been wanting to use them for months at the time, and I realized they just didn't work. So what you have to realize is that To get someone to switch away from Amazon or Google, you have to provide a significantly better experience in addition to a significantly better price. So what we— one of the core tenets of VectorDash is that, and I tell this to my 2 co-founders all the time, is that even if we were 10 times more expensive than Amazon or Google, people should still be willing to train their models on VectorDash just because it's such a pleasant experience. You have to focus on also being a really, really great product for people to use. Pricing aside, if people don't know how to use your product, they're not going to pay for it. And what we do is we talk to quite a bit of machine learning researchers and other developers, and we ask them, hey, would you use this feature? If they say no, we don't build it. Then we ask them what they want, and then we build it. So people have— let me give you an example. Someone recently asked to use multi-GPU instances rather. So instead of having one instance, instead of having one GPU per instance, they actually have one instance with like up to like 8 instances or 8 GPUs. We didn't do this at the time. We had one GPU per instance. You could spin up multiple instances, but it was one GPU per instance. So we decided to add multi-GPU instances, and we realized a lot of people want multi-GPU instances, which is pretty interesting. Also—

Very cool. It's not a shock. I mean, you can provision monster VPSs for when you're compiling code. So that makes sense to me from a cloud compute standpoint.

And also, it's really hard to sort of— sorry, I just lost my train of thought. So what we want to do is make it incredibly simple to train models, even compared to Amazon or Google. We want people to use VectorDash. Just because it's so much of a better experience. Just because we have— it's much easier to work with. There's less overhead in terms of configuring your instance and spinning it up. There's no network management. You don't have to set up a VPN or anything to connect to your instance. It's just click a button, press start, here's your instance login, you can use it. That's how simple it should be.

Very cool. So you want to become the Apple/iTunes of AI research.

Hopefully, yeah. That's a great analogy. Okay.

That makes sense. So again, kudos to you guys for being focused on all the right things as a startup. I really do think you're focused on all the right things. So audience, if you are wondering, am I a paid shill? Tell them, Sharif. You guys haven't paid me anything, have you?

No.

Okay, cool. And I don't accept bounties for these kinds of interviews. I think that as a community service, we just— hosts and the providers, They just need to be able to know that their questions will be asked. So that's why I'm taking live questions, and, and that'll be answered too. So appreciate you again being game for this, knowing that, that I wouldn't like float you a list of like, of like, you know, slow pitch questions like, this will be easy. Um, so thank you for doing this. But I think you guys are really— you're focused on all the right things. I really, I really think you are, as far as longevity and, and knowing that the project will have a future by serving a customer. You guys are doing great. So a couple of questions that have come up that I think are really important, and Crypto Mine Tech asks this one: can GPUs be undervolted at all? It's a very common practice in mining. Would that be okay? Have you done any of your tests with undervolted and overclocked cards?

Right, yeah. So, um, we have done some benchmarks on undervolted cards, um, and we found that for machine learning it's best to be at stock. Um, I'm sure overclocking won't hurt, but for the most part, just like GPUs are fine at stock. NVIDIA knows what they're doing. And undervolting, it does result in a pretty non-trivial performance decrease when training. So for the most part, just keep it at stock or overclock it if you're feeling fancy. Okay.

That makes sense. It seems like, I mean, CUDA certainly was— that was like all the CUDA libraries were designed with stock cards in mind. Yeah, it makes sense. Another question that comes up from LearnVFX, uh, aka my brother Aaron, who works in visual effects: what interface does VectorDash expose to the client, aka the AI researcher? So this is a security question, right?

Yeah, so we have, um, virtual machines on the client. So what we do— or for the host, um, so what we do is they're, they're 2 primary parts to VectorDash. There is sort of the central VectorDash server, which is where we coordinate pairing machines with users, and then there's the actual machine that runs the VectorDash client. What we do is we spin up a VM and then we just set up a proxy server that lets them connect to them, and they connect to the proxy server rather than directly to your network. So in terms of VM, we use the same virtualization AWS So it's fairly secure.

Got it. Okay. And let me just double-check for some other, some other questions. Xbitco is asking, how payout? How do you handle your payouts?

Yeah. So, um, we do payouts in 4 different ways. Um, you can, you can take your, uh, we do Bitcoin, Ethereum, and Litecoin, or if you want to get paid out in US dollars, we'll pay out via ACH.

Very cool. Yeah, very cool. Thank you for— I mean, oh my gosh, guys, what is that? Somebody refresh my memory. What was that movie? Do you accept cash? Yeah. ACH is great when it comes to alternate compute. It's just great, but the other crypto options are awesome too. What financial partners are you working with? In order to be able to make those payouts and be liquid? Because you're touching on an area of crypto, and as you scale up, financial partners, I'm sure you've been discovering, they can have mixed views on doing anything crypto-related.

Yeah. So we use Coinbase Pro. We pay out crypto via Coinbase Pro to your wallet. So there's this setting where you can just enter your wallet, whether it's Bitcoin, Litecoin, Ethereum, And then we can pay it out to you weekly. And then for US dollar ACH transactions, we're using Coinbase— or sorry, not Coinbase, we're using Braintree. So Braintree is a subsidiary of PayPal and they sort of do payments as a platform.

Got it. Okay. All right, guys, you heard it there. Let me see the follow-up questions. Follow-up questions from CryptoMindTech would be, are there— no, I think it was already answered about minimum number of GPUs to be considered.

Uh, no minimum, no minimum. You can have one GPU.

Yeah, yeah, awesome. And all right, and Robin is asking, is the code written in ternary or is it all in binary? Which Robin goes on to inform you will be less efficient in the future.

It's written in Python and then compiled down to Python bytecode, and then your machine runs it, your CPU runs it in binary. I mean, it's stored in binary. It's okay.

It's okay. I mean, again, we have a semi-technical audience and there are some members of the audience who are very technical and just know that their methodologies are right. So yeah, that's welcome back to mining. Yeah. Awesome. Guys, expect to hear that answer a lot, by the way, to anybody who's kind of more firmly planted in mining. Correct me if I'm wrong, Sharif, but Python is kind of like what makes the research world go round.

Right. Right.

Yeah.

So most machine learning research is— most development actually nowadays, or at least modern developments, either done in JavaScript or Python. And then I'm sure there's like high-performance libraries out there that are written in like C++. So, I'll give you like a quick overview of like how these machine learning libraries work. TensorFlow is Google's machine learning library. What they do is they've written the entire thing in C++, but they've written like a wrapper for that C++ library. So, this lets you use like Python's really, really simple syntax with C++'s really, really great performance. It's a great combination.

Yeah, very cool. So if they wanted to know more, like read a primer on TensorFlow then, sounds like, because you're focused on TensorFlow libraries.

Right. Yeah. So a great website just to learn about TensorFlow is tensorflow.org. If you read through the examples, they're really, really great. And even without a technical background, they seem to be pretty understandable, I'm willing to say. Yeah. TensorFlow.org, and then there is, if you really want to learn more about machine learning and actually start building projects, there's a great course called fast.ai. It's a free course. Hundreds of thousands of people have taken it in the past. I've taken it. It's awesome. You learn so many cool things, and it's, yeah, the projects are awesome.

Very cool. Okay, so fast.ai, and then guys, tensorflow.org. So I've got a couple other questions that I think are relevant, and then I don't want to keep you too long because you communicated to me that we can't go like 2 hours answering questions, and I respect that. I really appreciate the time that you've invested so far with the community to answer these questions to the best of your ability right now. So, and just so straightforwardly, thank you, especially when it comes to like the settlement portion, right? And payouts. That's huge to this community. So one person asks, can they, can they rent them out for mining, cloud mining?

You can.

I'm going to answer that one. Yeah, you think I'm going to answer, but I'm going to say to my community, please don't do that. These are the hardware requirements for, for what VectorDash needs are so high that you're— it's going to be onerous for the, uh, for the host to build up their systems. And only to have you turn around and mine on them. They can stay on their, their 1x riser cards and keep mining if that's what you want to do. So talk to somebody else in the, in the group about just renting their rig personally, if that's what you really want to do.

I will say, um, people have actually used SpectreDash instances for developing these mining algorithms. So if they don't have like a 1080 Ti, they can just rent an instance with a 1080 Ti and test their algorithm on it.

Okay. That is cool. Yeah.

For real.

For science.

Yeah, for science.

Okay. Yeah, cool. So somebody else is asking, all right, what's, what's keeping them from becoming centralized to maximize margin?

Um, that's a great question. I guess you could just, um, I mean, Amazon is centralized, Google is centralized. At the end of the day, um, we're a marketplace, so we have people willing to pay prices and then we have people we're willing to accept prices in return for their hardware. I guess this goes with any startup. Like, you could give the example of like Airbnb or eBay or like Uber. Being centralized does have its efficiencies in the short term, but in the long term, I'm not sure about that. So being centralized is much easier in that you can develop much easier. You can release software much easier. And it's much easier to develop a centralized platform for like this sort of distributed mining or distributed machine learning Then it would be to develop a decentralized one. So yeah, I mean, I don't know. I guess that's a great question. I'll have to think more about it.

Okay. Yeah. It's the whole like, yeah, Google starting with the mantra of don't be evil to where we are today.

That's true.

That's true. No comment. No value judgment over here, but some people have changed their opinions about whether or not they followed their own advice. So I don't know, guys. I mean, I think to speak on from a place of advocacy here, based on the conversation that we've had so far, I view what you're doing as far as there being a marketplace and brokering the deal between qualified buyers and qualified hosts as being one of the most valuable positions you can be in.

Sure.

You could probably find a large data center that's like, oh no, twiddling their thumbs, like, we've got all these GPUs, what are we going to do? You probably could find them. Yeah.

At the end of the day, like, the end goal is just to make AI research available to everyone. Like, I myself, like, I had, like, I'm using a laptop with like an Intel i5. And I remember like waiting like almost 2 days, like with it running nonstop, which is like trains, like recurrent neural network. And it was like, it was a fairly small set too. It was only like 20 megabytes. It was taking so long. So like, and I'm like a computer science student at like, a university, think about all the other people who are like across the world who like have access to the internet but not necessarily like a powerful computer. Um, I don't know, it's— the possibilities are endless.

Yeah, right? Somebody call Nicholas Negroponte, the, uh, the One Laptop Per Child. Be like, one laptop with VectorDash, the VectorDash one-click in the— be cool. So Possum Mining is asking about AMD GPUs and the prospect for having, uh, like what the outlook is And maybe even a timeline for having AMD GPUs rolled into the platform?

Right. Yeah. So the thing is, we can support AMD GPUs. It's just that most people don't use AMD GPUs for machine learning because the support for— they don't support CUDA. And TensorFlow, PyTorch, most if not all machine learning libraries are optimized for CUDA. It's not necessarily that it's hardware incompatible. It's just that the software ecosystem around AMD GPUs isn't as robust as it is for NVIDIA GPUs. So until AMD begins to step up their game, they begin to write, like, they begin to optimize machine learning frameworks for AMD GPUs, like, we can't really do much about it. It's up to people who actually use the GPUs to decide which ones they want to pick. And for the most part, they're picking NVIDIA right now.

Got it. On the demand side, have you had any prospects communicate to you that they would be interested in buying AMD GPU compute power?

Um, we have not. No, unfortunately not.

So you're saying there's a chance?

There's a chance. Definitely a chance. Um, but right now most people are pretty keen on using NVIDIA GPUs for machine learning.

Makes sense. Okay. Yeah. Sorry guys. Womp womp. Bad news. But, um, but it was, it was framed as accurately as it can be.

Go email AMD and ask them to write better drivers.

Yeah, I've been there, done that on the mining side. Um, I've been doing that forever. Um, awesome. I got a question also here, and it's kind of a rehashing of the question that— sorry, that joke comes up a lot, that pun comes up a lot— rehashing the question that I, I had before. Is it possible to use— wait, no. Um, hmm. So somebody's asking about, about clusters. I think that's going to be a recurring question from my community because we're going to have A— we're gonna have a multitude of small miners that have— like I told you a little bit about the hardware that I have just here in the office, right, for, for testing algos and for doing vBIOS, uh, flashing and stuff. And so a small system like this with a— gosh, I guess maybe 3, 4 motherboards where I could build up, um, I could build up some of them and I don't know, maybe get that, that 5th motherboard here in the office. with a Threadripper so that I have a higher core count or a dual Xeon system or something that then can be my controller node for that cluster, that compute cluster, and have it fit.

Right. So— Yeah.

What are the prospects there?

That's a great idea. The thing is, like, the way the bottleneck works is you're only as fast as your slowest link. So even if you have like a node with like dual Xeon CPUs, If the machine's too— like if the GPU on that PCIe slot's still like a 1x slot, the link will only be a 1x link. So you sort of have to take the minimum or like the slowest portion of that entire connection and that'll be the entire connection's bandwidth. So it's like in theory it's a great idea, but in practice it just doesn't work out that way.

Sure.

The best way to do it would just be to get like a motherboard that can support like multiple 8x or 16x slots. Um, I think there are a few. I think Octominer might have one. I'm not sure. I'll have to double-check.

I think their next-gen— they're, they're working on an 8x, but I don't think that that's what's broadly available right now. And Octominer is not necessarily doing the R&D on that. They are— they're buying from other Chinese resellers just like the rest of us.

Interesting. I didn't know that.

Yeah, so I've got a— I, I often talk about— I've— and guys, just— there you go, disclosure. Um, I've got my own little line here, commercial line called Big Baby. Um, and it's— yeah, it's the Hellboy reference, but it's called Big Baby. And I, I know which— I, I know where all the silicon comes from. I know where it's all put together, and it's put together by the same folks. So yeah, 8x multiple riserless boards are— they're not here yet. I can tell you that safely. They're not here yet. So that's why we ask about alternate ways of setting it up. As far as there being like appropriate network fabric, are you saying then we'd have to have like what, like beyond 10 gig E network bandwidth, right? Like well beyond it to start approaching what's needed, right?

Right. Yeah. So too, if you were to have multiple Like, if you ever have like a central CPU, like with duals, like a lot of RAM. Like, even with the RAM, like if you have like 8 gigs on like the GPU worker node and like 64 on the central, like the main node, like the RAM is like, it's, you're only as fast as your slowest link. So you'll have to go over like InfiniBand or something like that. And it's just not going to be fast enough.

Got it.

Yeah. And like, got it. The, the, like the amount of money you'd have to spend to like purchase the hardware necessary to make it fast enough would be much better spent just to like purchase like separate machines.

Got it. Okay. That makes sense. So, all right guys, you heard it there. We've asked a couple of times in a couple of different ways about, about setting up clusters. And it sounds like if we can find a qualifying motherboard that just helps resolve some of the, the total system bandwidth problems, then our CapEx is better served that way. Somebody's saying, unfortunately, all GPUs are currently being used. I think somebody wants to buy VectorDash instances. They want to buy VectorDash GPU compute, and they can't buy any.

So we're— yeah, so we're actually at capacity right now. But we are bringing more machines online this weekend. So in about 2 days, I believe, we'll have more online for use. So it's temporary while we're spinning up more GPUs to use.

Nice. And yeah, we're getting, we're getting, um, getting some people saying like, wow, I really like this project. And somebody else said it's the Seth effect, he'll make you like every project that comes on his channel.

That's awesome.

So, well, let me make sure that there's any other questions that are, that are relevant, and I want to let you go in, in a timely fashion. Sorry that I haven't been able to multitask and read all of this stuff. XBitco saying, I'm waiting for Seth to test VectorDash first. Okay, Sharif, can we, can we set up a test? Is that possible? I've got, um, I've got a qualifying system over here.

Um, definitely.

Yeah. Okay. And then do the hands-on. Is that all right as a follow-up?

Sure thing. Yep.

Okay, cool. Yeah. Green light me. I've got the Core i7 over here with some, uh, some 1080 Tis.

Nice. We do need more 1080 Tis. Uh, lots of people like 1080 Tis. Okay, cool.

So I'll create a dual boot image over here for Ubuntu and make sure there's enough hard drive space for that.

Okay.

So any other questions right here? I've got another one. Harmony is asking, when coin? They're expecting you to launch a token.

So there is no token. There is no ICO. Unfortunately not. But if you want to rent out your GPUs, fill out the hosting form, and then we'll get back to you. Awesome.

And LearnVFX asks again, how about generalizing the Marketplace? If it works for TensorFlow, can it just work for any GPGPU workload I want?

So we do expose, like, you can use an entire Ubuntu VM. It's just that TensorFlow comes pre-installed. So you're free to install whatever you want inside the VM. Like, you can run a mining algorithm in your VM and just mine away. I don't know why anyone would do that, but it's a VM. You can do whatever you want with it. But most people use TensorFlow and PyTorch just because it's what's most economically feasible right now.

Awesome. Would you be opposed to that? Would you view that as scope creep for the VectorDash project if on the demand side There are requests for things other than TensorFlow and other AI libraries?

Not really, because we give you access to the full VM, so you're free to install whatever you like. So I think if you want to install a different library, like for example PyTorch, it's one command, it's apt install PyTorch, and you have it installed and you can use it.

Okay, awesome. So essentially you're not really limiting the demand for anything else anyway, and you're not really monitoring what it's being used for.

Right, yeah. So you're free to use it for whatever you want. whatever you want since it's a VM. So most people seem to prefer TensorFlow and PyTorch and other machine learning libraries. But if you want to render a Blender file using the command line, you can do that too. We're not stopping you. You're free to use it for whatever you like.

Okay. Transcoding video material, whatever.

Yeah.

Yeah.

We do have people using it to encode videos too.

Very cool. Very cool. All right. That's good to know. And all right, a couple other questions and I think we should probably, should probably wrap it up. What do you say?

Yeah, sure thing.

2 more quick questions? Okay. So, uh, Abdul asks, how long should we expect to, uh, to wait before receiving a reply, uh, if they've like, if they've submitted their info to be considered?

Yeah, so what it—

Because it sounds like you're not ruling people out yet, you're more trying to open To as many people as you can?

Yeah, so right now we're just looking for machines that we have that we need to use for development. So right now we're still in a private beta. What we want to do is, like, currently as we speak, we're looking for machines that match what we want to test on. So if we have, like, this new AMD library and we want to test on AMD CPUs, we'll go out and reach out to hosts who have those AMD CPUs, which is why we ask what CPU type you have. In terms of general, like, opening up to the general public, within the next couple of weeks, if not 1 month or 2 months. We're trying to work— we're trying to develop that it works on a wide variety of hardware, on heterogeneous hardware, because it's really easy to make it work for like 1 type of CPU, 1 type of GPU, 1 type of motherboard, but making it work across all different types of CPUs, GPUs, and motherboards, like the possible combinations are like nearly endless. It's crazy. And we just want to make sure the people who are actually paying for the instances are getting decent performance out of them, and the people who are actually using them are using a secure and safe client. We don't want any privilege escalation or VM breakage, anything like that, which is why we're taking our time with developing it.

Okay. Yeah, that's great. Great to hear. Another question, and I think we'll probably wrap with this, guys. But the question is managing expectations for the demand side, for clients. What kind of uptime are you guaranteeing? Are you giving any kind of guarantees for uptime since you're outsourcing the hosting?

Yeah. So the way it works is that you just set an availability. And so for example, if you have like a rig and you want to list it for a week and you're confident you can keep it online for that duration of a week, you just set the end date for 1 week from now. And then let guests see that, hey, this machine's online for a week. This model I'm willing to train is only going to take a day or so. I can probably use this machine. And so as long as you keep your machine online for that duration using the instance, that's fine. On the off chance you have your machine go offline before the availability end that you specified, the instance is free for the end user. So they don't have to pay for anything because the machine went offline.

Okay. Awesome. So you've managed expectations pretty well then on the client side.

Right.

Yeah. On the buyer side. Okay. Very cool. One last question that I think is actually really important, and shout out to Arstein for asking this. Should the cards on the host side, should they be the same model, brand, specs for right now? You mentioned trying to code towards proper heterogeneous environments, but today in participating, To be considered more seriously as a candidate, should they have all the same model GPU?

So the actual manufacturer model doesn't matter, but the— it's like the 1080 Ti or like 1060. Um, because we let people spin up multi-GPU instances, um, the GPUs on that instance have to be of the same type. So we can't multi-GPU instance with a 1060 and a 1080. It has to be 2 1080s or 2 1060s. So if you do have different versions of GPUs, what I recommend doing is keeping all the 1080s on one machine, all the 1060s on another machine, all the 1070s on another machine, just because if anyone wants to use a multi-GPU instance, they'll be able to use those 3 1080s you have on that same rig versus that one 1080 if it were like with a, with the 1060 or 1070. But like the actual manufacturer itself, like whether it's like an ASUS or like a Gigabyte GPU, it doesn't matter. Like the model matters Got it.

Okay, so, so specs then. All right, the last question. Wow, the guys— guys, why didn't you put these questions in the comment section before? Um, because some of these are actually really good. CryptoMindTech is asking, does the rig rebooting constitute going offline?

Uh, rebooting does constitute going offline, um, but we've made it so that it's only if the user complains about the machine going offline that there's an issue. I'll explain this a little bit more in depth. If you shut down your machine with a VM active on it, and then it reboots back up in less than a minute, the VM will still be there, the connection will still persist, and the workload will still be running. But if you have a machine go offline due to a power outage or something, or there is a hardware failure, and it's offline for hours or much more than the end paying user would be willing to tolerate, then that's considered offline. But for the most part, going offline for even just like a minute is considered going offline.

Got it. Okay. That stands to reason. It seems like with mining in particular, it's not processes that are extremely— the process of hashing is fairly antifragile, we'll say. So you can have multiple reboots in a day and it's like, oh no, you haven't lost all your work.

Yeah, there's quite a bit of state to manage compared to mining with VectorDash.

Okay, awesome. Guys, smash that like button. Make sure you go over to vectordash.com. Check out the hosting page right here. I've got all of the profitability right there. Can I ask a couple of project questions that are— this is just for posterity's sake. 5 years from now, and you know how startups go, 5 years is forever. There can be a lot of uncertainty with startups. Let's just pretend like something catastrophic happens. Knock on wood. You know, heaven forbid that it happens to your project, but pretend like something catastrophic happens and Vector Dash is no more 5 years from now. What do you want people to look back on and remember from the team, the project as it stands right now, or your work to this point?

Yeah, so I guess if I can say that I moved the needle of progress just forward, just even a little bit. I'd say it was a success. Um, like, the end goal with VectorDash is to be AI. Um, that's what we want to do. So even if I somehow figure out a way to move it, just if I figure out how to like make us go a little bit faster, I'd consider that a success.

Awesome. Well, I know you put a lot of hope into miners' hearts for sure, so Yeah, again, knocking on wood that you guys do, do go well past your, your first 5 years in business. So thank you for making time for the show. Appreciate you doing this. And, and normally I ask a bunch of like silly questions about like, hey, we're going to pay me that $25,000 you owed me for this interview, but we already covered that. So thunder stolen. So appreciate your time. Thanks so much for making this happen. Uh, where can people get a hold of the VectorDash team if they want to reach out to you personally? Is there high availability to talk to team members? I know you guys have coursework to get to also for school.

Right. Yeah. So if you actually click on that little box in the bottom of the page, that goes directly to my phone. So whenever someone messages that, I usually respond within like a minute or so. Yeah. Yeah. It goes directly to my phone. So if I'm in bed, I get like a message, I'll try and reply to it before I go to bed. 2 AM and like, yeah, which is when I'm sleeping. Um, but yeah, if you— you can use that chat box in the corner, or you can use— you can just message me on Twitter, or you can send us an email at team@factor-dash.com. I just got a message.

Cool. Is there a Discord group or a Telegram group, or, or is there a different like forum that people can, can congregate in for the Vector Dash?

So we don't have a Discord or a Telegram, um, but if you do want to keep up with updates, um, you can follow me on Twitter. I usually tweet about Vector Dash pretty often. Um, my is @samain100. Um, or you can follow, uh, Vector Dash GPU On Twitter. But for the most part, just create an account. We'll email you updates once we're ready to bring on more hosts. And yeah, I guess that's about it.

Awesome. Sharif, thanks so much. Really appreciate your time. Really appreciate you giving the community this very, this very close look at the project and answering so many questions. Really appreciate it, man.

Thank you. Yeah, thanks for having me on, Seth.

All right, guys, we'll see you soon. Bye.

Don't fall. Concentrate on the figure or you will lose all that heavenly glory.

Yes!