WEKA and Micron leading the way in AI, hybrid cloud, and sustainable data solutions
[Music] thank you for joining us today I'm Jeremy Werner I'm the corporate vice president and general manager of Micron storage business unit I have a fantastic guest with me here today Liran Zvibel, who is the Co-founder & CEO of WEKA welcome thank you Jeremy for having me Liran and I have known each other for some time and WEKA has really developed and blossomed with the takeoff in AI as a leader in data platform technology maybe you could tell us a little bit about what that journey was like for you what was the genesis of WEKA so WEKA is actually our second storage company after we were done with the first one IBM acquired us we got a lot of experience we identified two main issues with the legacy storage market one the market is insanely fragmented there are dozens of products that actually sell and customers have to go and choose the right product for the right application sometimes customers have 40 different products running in their organizations the other problem was that even though the engineering is all software 99% of the engineers are software engineers the delivery is always in some proprietary hardware box it makes it very difficult on-prem when you're buying your service from one vendor and the storage from another vendor but it makes it impossible to run these on the cloud okay so we didn't want to do that anymore we said never again then in 2013 before we decided to found WEKA three major innovations happened microservices with containers happened NVMe happened so finally you could connect flash directly to the CPU with no controllers third networking has really caught up to compute speed so traditionally the network was way slower than the server nowadays it's as fast or even faster taking the three components together instead of taking the name space and breaking it down to controllers which where it controller owns a smaller portion of the NAND space like a traditional storage array then if the clients choose to all access whats on a single controller they're basically doing a denial of service attack to that machine what we have done we're taking all of the data all of the metadata breaking it down to tiny little pieces and then we're running a perfect load balancing of all of the iOS data iOS through all of the NVMe devices and all the metadata operation through all of the CPUs so we were able to create a system that runs all of the workloads better than the one-off products and doesn't require any specific hardware so it runs on premises and the cloud wow so what you end up with is a very flexible scalable and high performance data platform for to be able to serve all applications in a data center but especially you've had a lot of success lately as AI's taken off in the data center so huge congratulations to you you recently announced your series E funding $140 million at a $1.6 billion valuation it's an amazing accomplishment I personally was involved with three startups none of them ever got to unicorn status so I know how much effort goes into achieving such a huge valuation Micron was an early investor in WEKA so I know we always believed in your capabilities as a company recently I saw that NVIDIA was a big investor in the series E now I think that has something to do with your success in driving AI platforms tell us a little bit about this series E what it means to you and why NVIDIA would participate in this round yeah it's a great question and thank you for the congratulations it's a very big milestone for us so from the bird's eye when you ask a VC hey what would make a company a startup successful you need two components you need the market to take off and you need the 10x advantage and fortunately we have both when you're looking at the data center and the new projects they're they're built around this triangle of compute network and data NVIDIA was able to get compute and network to be four five six orders of magnitude faster than what was possible just a decade ago with standard CPUs and previous kind of ethernet when you have made compute and the network so much faster you also need the data to be so much faster when what we're showing is that we can fully saturate all the links we provide a lot of more small iOS and it works well with a smaller file so when customers switch to us they're actually getting much better utilization out of their GPUs so when customers tend to WEKA when they're comparing hey I can buy products that were built on ideas that made sense in the '80s and the '90s and the 2Ks that's 100% of our competitors there not realizing their full potential when they're buying new kind of computer new kind of network they also need to buy new kind of a data platform so this AI explosion is just a huge opportunity for your data platform solution for sure and another part that's interesting for AI and we're probably going to touch on it on another question is AI is not an application it's not a code anymore it's a pipeline and through through that pipeline you're starting with ingesting you're running the MLOps or LLM Ops ETL you're training your checkpointing you're then running a huge regression making sure the new model is better you're inferencing in production and you're archiving you have so many more steps if you're if you're subscribing to the old mode of hey each step has a different system you're wasting a lot of resources and you're not making as much progress as you could and I think all this training's come a long way with LLMs with picture generation image generation models like stable diffusion but there's still a lot of opportunity ahead of us in terms of what's possible with AI we barely scratched the surface I think on training with video for instance for sure so we're seeing generative AI really starting to take off but the way we're we're experiencing the market now we're we we're working with a lot of startups building technologies we have not yet seen the large enterprises really buying them really leveraging them enterprises are still in this discretionary mode where they're running small POC's where we think the market's going to go again predicting is hard especially the future we think that in the next two to three years we're going to have a lot less of these startups maybe the best 20% are going to remain there but now we're going to find out the best ideas for leveraging AI and generative AI in enterprises and then the big companies the corporations will really start leveraging these for the day-to-day operations we we're looking at that mode like mobile was in ' 07 you know when Steve Jobs launched the iPhone I think a lot of people realized hey there is a huge potential there but it took about five more years for different kind of apps for people to fully comprehend hey this thing the fact that they have another computer in my pocket changes everything AI and Genie are closer to the iPhone of '07 than to the iPhone 24 yeah we're truly at the very early stage of this transformative technology for sure I really believe that simplicity is key and you talked about the fragmentation of storage systems I see it when I talk to customers the it's really hard to find these system administrators who are able to manage so many different storage solutions converge solutions network solutions different object file block so being able to provide that simplicity absolutely has got to be a huge value for your customers or our mutual customers I actually know for a fact one of the simplifications that I think customers are looking for today is around the hybrid cloud now you and I had spoken a couple years back about the hybrid cloud how it was going to be something that all enterprises would embrace because there are still some benefits of being on-prem but also benefits of running some of your workload on the cloud where are you in that journey where are your customers now in terms of adopting hybrid cloud and what kind of solutions are you offering so you're you're spot on hybrid cloud is something that we used to talk about in future tense a a few years ago but now it's really starting to happen I think the were were many organization that were religious on staying only on premises there were some that said no we're all in on the cloud I think both realize they're not getting the best volume the cloud is best utilized for elasticity if you're just taking everything you've run on-prem you're transforming it one to one to the cloud it's going to be really expensive but if you're not leveraging cloud at all for elasticity you really have to provision on premises data centers that are way bigger than what you need for a majority of the time the smart customers we're starting to see are leveraging on-prem for 80-90% of the time we're running most of their workloads and then they're leveraging cloud either for DR or for bursting of added compute power when they need that some of them store the capacity on the cloud as as well but some actually choose to build a a flash based object storage in some Colo and equinix or similar that they're controlling on premises they can extend the name space from their on-prem places to their they can also extend the NAND space from the cloud to there and they can really manage the costs in an impressive way that's so cool so you know that is a vision I've always had for companies because obviously like if you're an eCommerce company "Here Comes Black Friday" your compute needs are going to spike through the roof but then you know other part times of the year you can satisfy what you need on-prem very cost effectively so one of the things we've seen a lot from our mutual customers is the adoption of the Micron 6500 ION 30TB drive now this product has won a lot of awards over the past year why do you think this product in particular has become so popular with your customers so I'd like to start by saying that I think the ION 6500s are one of the best engineered flash products now on the market thank you so much you were able to cram a lot of capacity 30 terabytes come up with a QLC like price point with a TLC like performance so engineering is all about making the right trade-offs making the right choices and you have made great great choices in in making this device we have basically two different users to flash in in these AI GPU accelerate workloads there is the pinnacle of performance a small portion of the capacity that needs insanely high amount of iOS and throughput and this is where the 94-9500s work out if in large capacities maybe you can get it with the 7450 but then most of the capacity that we're seeing for these gen AI projects are actually in the overall corpus where you still need low latency you still need good amount of throughputs we're now seeing project with hundreds of petabytes we're now seeing several project that are exabyte scale you know a year ago had you told me hey Liran you're going to have customers buying an exabyte or two or now we have five in a single project they would have told you Jeremyy you're out of your mind like looking forward is good but that's never going to happen okay I'm going to tell you today two years from now we're going to be back sitting at this table and we're going to be talking about 10 exabyte deployments in the data center running on your software I by the way now I can see it happening more than two years ago I would have seen an exite now I believe it's amazing how fast the needs for relatively fast but also cost effective storage have grown one of the things that's driving that need and the TCO equation from hard drives and many hard drives you know scaled out to consolidated racks of solid state storage like what we're providing is sustainability and what I one of the driving impetus for sustainability is the fact that we're running out of power to in this country and in many countries around the world in fact recently there was an article by the Washington Post that reported the same that America's running out of power and I saw a statement that stuck with me the other day which was there is no backup planet which really made me think about some pride in terms of what we're doing um to be able to bring the power consumption of the data center down with solid state storage maybe you could talk a little bit about how SSDs and the data platform deliver better power utilization in the data center and better efficiency at the data center level yeah so just to begin with I concur completely we do have a problem now our last round was led by Valor the previous round was led by Generation Invested Management it's a fund founded by Al Gore the form of VP and David blood they care mostly about sustainability they invested in us because they have figured out hey we're we're making a big dent in how can you save power also what we've learned from them you know a decade ago data centers consume one and a half% of total power consumption nowadays it's 3% in a couple of years it's going to be 8% wow so an exponential growth and at that point data centers are going to consume more power than all of the human workforce combined so we do have a problem what we're seeing with our customers when they pick WEKA two things happen one because we provide data so much quicker to the GPUs we are getting their utilization to be way higher so one customer that adopted us on a WS instead of FSX for luster at very large scale started with a problem that their GPUs were only working at 30% utilization throughout throughout the day when they switch to WEKA we're able to show them that they're getting six times the throughput and 40 times the iOS but data scientists don't get don't care about throughput and iOS what they care about is getting their job done what they realized hey we got their GPU utilization going from the 30% a day to the high 80s so the same GPU cluster was able to deliver three times the output so if you can generate three times the output using the same amount of power it's a huge saving sure the other part where we make a huge difference because we support all the protocols because we support all of the work clouds you don't need to start copying data around different systems which happens today you injest to the wacka you run your envelops ETL portions on the WEKA you run the training the checkpointing on the WEKA you run the large scale validation on the WEKA the inference on the WEKA and archiving also happens on the WEKA you're saving the job of keeping three or four or five different platforms and copying data around this is also a huge savings we''ve commissioned some work to to try and quantify it and they found out just on the effect on the storage when you're using WEKA a petabyte of WEKA saves 260 kiloton of carbon dioxide emission per year which is a huge amount that's amazing so helping the planet saving customers money making data centers more efficient improving the efficiency of the entire infrastructure and bringing new capabilities to humanity thank you so much for being here today truly a pleasure to have seen you and your vision come to reality and everything that you're doing for WEKA and for the industry in general just want to thank you again for being here yeah thank you very much Jeremy for having me I really like having these discussions so thanks and thank you for joining us we hope that you enjoyed our discussion today and learn more about how WEKA and Micron are collaborating to bring you some of the world's most advanced Storage Solutions [Music]
2024-08-25 15:19
