The World's Largest AI Supercomputer (36 ExaFlops)

now one of the biggest problems that new AI Hardware vendors have is who exactly are the customers all of these brand new companies and I'm tracking dozens of them that have tens hundreds of millions of dollars of funding to build new chips for machine learning they need to go out there and they need to see where their revenue streams come from big number think let's go sell to the hyperscalers let's sell 100 million units and we can be the next Nvidia however reality is a little different now I'm holding this this is a cerebrus wafer scale engine or at least you a metallic version of it and today's news is that they've got a customer who's basically pulling an order in for about a billion dollars of this stuff foreign [Music] a lot of the content on this channel wouldn't be possible without you the supporters many thanks to all who support and you know if you're interested in supporting then we have patreon we have a merch store I have a substack newsletter or simply just like And subscribe it really does help out the channel so I've mentioned cerebrus on this channel before but if you've never heard of them let me bring you up to speed this is one chip and this is real size it is essentially wafer scale and they call it the wafer scale engine 46 225 square millimeters 2.7 trillion transistors 850 000 AI cores and 80 gigabytes of on-board SRAM 100 yield here because standard wafer has about 50 defects they have so many AI cores that they're built such that they can just be routed around the ones that have the defects one full chip at least a CSR two the second generation is 24 kilowatts and for Price well we do know that Pittsburgh super Computing Center bought two for five million dollars so that puts the upper limit of the first gen at two and a half million second gen who knows maybe it's arm leg and firstborn as well so what is the news today well these like I said these AI companies have to know where the revenue is coming from and something as big as wafer scale you're thinking how many people in the world actually need one of these and why do they exist in the first place well why do they exist in the first place you've seen that there are companies buying up hundreds of thousands of gpus to build their machine learning models why because they can't fit the models on one chip you have to as when your model gets large enough go into two chips four chips eight chips and that has inherent latency uh problem Associated you also have to build your software to be able to send the data around relevant to each of the batches of the machine learning code and that takes memory bandwidth and power and latency as well if you you can have fewer bigger chips and that solves a lot of those issues that's why the cerebris wafer scale exists now cerebris's customers that they've mentioned today are people like GlaxoSmithKline oil and gas HPC and they've even found some use cases in defense but beyond that for a company that's selling something this big for a couple of million a pop you've got to think where are the customers now cerebrus reached out to me saying that they've got a new announcement with a large customer buying a shed load of these things and initially I thought well that's got to be one of the big hyperscalers the Amazons the Googles Alibaba Baidu 10 cents something like that but then I realized if that was the news then that could be the death knell for the company because it'd mean that they would only have essentially seven big customers available in the world well today's announcement is that a tier 2 cloud provider in the Middle East called g42 is set to buy or is at least penned to buy over 500 of these how do we get to 500 of these well it all starts with something called The Condor Galaxy this is going to be the minimum irreducible unit of 64 chips built with cerebris wafer scale engine twos connected with cerebrus's swarm X technology the first system is currently has 32 of these chips and will be extended to 64 in the next couple of months that is going to be based in the Bay Area and controlled by cerebrus in cerebrus Cloud but it's also being housed at the same facility that cerebrus's Andromeda system the 16 wafer scale engine is available to customers in the cloud already there are two more 64 wafer scale engine-based systems two more Condor galaxies going to be built in the US over the next 12 months one in Austin and one in North Carolina these three are going to be networked together to form one massive machine learning Beast even with the additional latency of going from state to state they believe that it's still going to be a Powerhouse when it comes to large models especially when it comes to machine learning with llms or other big models that need lots and lots of data now that's just three systems there are going to be nine here where are the other six going to be well the the Agreements are essentially in place essentially just need to be executed on based on delivery of the first third uh but systems four five six seven eight and nine are going to be outside of the US uh exact location TBD but again it's all going to be as part of one big worldwide Network it's actually going to be managed by cerebrus and any time that this first major tier 2 cloud provider client doesn't need the systems that time will be sold on Sirius Cloud kind of like Andromeda is today so who is this big customer well as I mentioned a tier 2 cloud provider this one based in the Middle East in the Emirates called g42 it's one of these a big conglomerate movie dialer type operations and as one of the biggest cloud provide providers in the area they obviously have a dedication to building the next generation of AI for their customers and they've decided one of their angles is to go down with a lot of wafer scale engines and one of the big things that g42 does is what they call m42 Health so this is where you have lots and lots of intrinsic Health Data and try to draw parallels to help the predictive nature of that data in the future now the Emirates health system and m42 itself has access to lots and lots of health records and obviously you have to factor in all the data protection and such but the idea is this is an opportunity to be the build a model around the biggest collection of Health Data and process it at scale without having to wait for a hundred and fifty two hundred thousand gpus or a billion dollars of funding instead it's going to be a billion dollars of these now let me tell you how I get to that number so we have 64 Condor Galaxy systems at 64 chips per Condor Galaxy and they're going to nine Condor galaxies that's what 560 570 chips now we've been told by cerebrus that one of these 64 chip systems is going to be around about 100 Mil plus minus Modified by nine and a billion isn't far off especially when you're having say support contracts and everything else perhaps uh well this does mean for a company like cerebrus who has about 700 million dollars in veg Capital funding they are able to demonstrate that they this through Just One customer they have Justified all of their VC funding to date this means that they can go after other tier 2 Cloud providers elsewhere in the world there are plenty of them out there and they're not even scratching the surface with any of the tier ones any of the super seven so this means that there's still a long road ahead for cerebris there's still a large Marketplace for chips of that size of this caliber and where the money for the next batch of Revenue could be uh people always say well the goal of these companies is either to be acquired or to be publicly traded there's still some pretty significant companies that aren't publicly traded out there um so cerebus at least for now it looks like has secured its future we're just waiting on wafer scale engine 3. Andrew come on you know you want it so thanks everyone for watching I'm gonna see if cerebrus will actually let me make these and then perhaps we can sell them in the store as you may or may not know I already have a mug with a cerebrus wave scale on it you can buy in uh in in the merch store very reasonably priced and I do have somebody on Twitter already saying that they're looking to put in a big order so if you're interested in one then go check it out but if you're interested in one of these also let me know let me see what I can do foreign [Music] foreign

Channel: TechTechPotato

Views: 15,474

Keywords: Cerebras Systems, Wafer Scale Engine, G42, M42 Health, WSE-2, CS-2, Andrew Feldman, Machine Learning, AI Supercomputer, Exaflops, TechTechPotato

Id: oVHkXEzKzxM

Length: 9min 18sec (558 seconds)

Published: Fri Jul 21 2023