SC25 | Garima Kapoor, MinIO & Avi Shetty, Solidigm

Clips
More from SC25

Garima Kapoor

Co-Founder & Co-CEO

Minio

Avi Shetty

Sr. Director—AI Enablement & Partnerships

Solidigm

play_circle_outline New Lab Launches SSD Benchmarking and AI Infrastructure Enhancements: Unpacking Interdependent Factors in Performance Efficiency

play_circle_outline Solidigm showcases cutting-edge SSDs and high-performance networking for AI applications.

play_circle_outline Enhancing Hyperscaler Training and Inference: Collaboration Between Solidigm, MinIO, Intel, and Supermicro in ExaPOD Reference Architecture

play_circle_outline Customers eager to benchmark SSD solutions and optimize deployment strategies.

Info
Transcript

Garima Kapoor, MinIO & Avi Shetty, Solidigm

Garima Kapoor

Co-Founder & Co-CEO Minio

Avi Shetty

Sr. Director—AI Enablement & Partnerships Solidigm

In this interview from SC25, Garima Kapoor from MinIO and Avi Shetty from Solidigm join theCUBE’s Dave Vellante and Savannah Peterson to discuss the "three horsemen" of AI infrastructure: compute, networking and storage. The group details the launch of a new AI lab designed to move beyond standard datasheet benchmarking and provide real-world performance data for complex AI dependencies. Shetty highlights the lab's cutting-edge specs, including high-density 122TB SSDs and 800-gigabit networking, while Kapoor explains how MinIO’s software-defined storage maxim... Read more

explore Keep Exploring

What changes have occurred in SSD benchmarking and what factors are now considered important in the AI context? add

What was the goal of investing in the AI lab and what announcements were made regarding its capabilities and partnerships? add

What are the specific workload requirements and considerations for hyperscalers when building their environments for training-driven tasks? add

What are the needs and considerations of OEM customers and end customers regarding SSD solutions and performance benchmarks? add

bolt Powered by CUBE AI

Garima Kapoor, MinIO & Avi Shetty, Solidigm

search

Savannah Peterson

>> Good evening nerd fam and welcome back to St. Louis, Missouri. We're here closing out day one of our three days of coverage on theCUBE at Supercomputing 2025. My name is Savannah Peterson, bringing you some extraordinarily, practical real- life examples of AI and HPC in the wild with Dave Vellante all week. Dave, we've had a really dynamic day.

Dave Vellante

>> That's actually, like you said, we're into the evening now.

Savannah Peterson

>> Yes, yes. It's dark out there.

>> Some nights are getting longer.

Savannah Peterson

>> Yeah. Everyone's favorite season.

Dave Vellante

>> Thankfully, we have some incredibly bright lights on the stage no matter the time of day. Garima and Avi, thank you so much for being here.

>> Guys, thank you for having us.

Avi Shetty

>> Thank you for having us. - I was joking, Avi, we're going

Savannah Peterson

>> to have to start putting you on this side of the fence.

Garima Kapoor

>> You've just been rocking it. I like this though. You brought us a lot of different dynamic conversations today with different folks in your network. I want to start off by talking about the new lab that you launched. Tell us about that.

Avi Shetty

>> Yeah, thank you Savannah for asking us that one, because we've had a fundamental shift in the way we do SSD benchmarking. Like earlier years, the focus was random reads, random writes kind the standard data sheet metrics at different queue depth. But in this AI world, there is a lot of dependencies. There are three horsemen, right? You have the compute, the network, and the storage. And all of them are contributing to the performance efficiency, the efficiency of the data center and as well as the bottlenecks. The bottlenecks are shifting based on you fix one, it moves to the other. And our goal with investing in this AI lab was to have us, our engineers, as well as our partners access to a real world AI infrastructure with the state-of-the- art GPUs from NVIDIA and the state-of-the-art networking, 800 gigabit networking where the GPUs are talking amongst each other at 800 gigabit. The storage is talking to the GPUs at 800 gigabit with Solidigm SSDs. And then with our partners like MinIO will manage that cluster for us. So as part of the lab, we announced the lab a month ago in October.

Savannah Peterson

>> Congratulations. - Thank you.

Avi Shetty

>> And we did two big cluster announcements. We announced the world's highest performing test cluster with our PS1010, Gen5 TLC SSD, and also the world's most dense test clusters, the 122 terabyte drives, which I showed you this morning. We have 192 of those in 16 1U slots getting you 23 petabytes. And we have MinIO software now managing that, thus giving MinIO and representation of actual TCO as well as our joint customers access to the lab where they can test out this infrastructure before they make their choices.

Savannah Peterson

>> And Garima, why is this testing ... And I have some hunches, but I'm curious, but why is this testing so important, and why did you want to make sure you were involved with Avi and the team?

Garima Kapoor

>> Yeah, I think one of the things that enterprises are struggling right now, there's a lot of excitement around AI and applications that you can build around it and the utilization of the GPUs. But what our customers are looking for is a very actionable plan. And this testing and reference architecture gets exactly that in terms of what kind of networking you need and what kind of compute do you need, what kind of performance you can expect to the T, And that makes it very actionable for our end customers to use it. So that why partnering with Solidigm was extremely important for us. First of all, the scale of the data, as we all know is exploding. Exabytes are the new petabytes, if I may say so, and there are multiple customers who are heading in that direction. They are growing from 100 petabytes to exabyte in a year's timeframe. And that's the explosion that we are seeing.

>> Wild. - It's insane.

Savannah Peterson

>> Yeah. - Nobody could have predict that.

Garima Kapoor

>> Even as you say that, even knowing this stuff,

Savannah Peterson

>> I feel my mind just boggled by the fact that,

Garima Kapoor

>> that's the scale we're operating at.

Garima Kapoor

>> Absolutely.

Savannah Peterson

>> And for that, you want a system in place that is scalable, that works on the TCO for the customers, and that gives them the performance that they are expecting for their AI application workloads. And that's where this lab is extremely important or exciting for us to partner with Solidigm, along with the Intel and Supermicro to make sure the customers get what they expect. I think that as much as you can narrow the gap between expectations and reality, that's where I think the success lies for the customers and that's what this lab seemed to deliver.

>> So ExaPOD was born out of this relationship,

Dave Vellante

>> which is a reference architecture.

Garima Kapoor

>> Yes. - What was the target for that? Is this for labs?

Dave Vellante

>> Is this for intense AI training? Can you describe that?

Garima Kapoor

>> Yeah. A lot of the hyperscalers, they're moving towards ... Hyperscalers have very specific workload requirement, which is training-driven. And for them, the scale is exascale. And for them, there are a lot of considerations that need to go in, this power consumption consideration. There is consideration in terms of the actual rack and stack, the data center footprint. All of them are considered when they're building out that environment. And I think as much as dense we can go on the drives that gives them a better TCO per terabyte. And for us, the analysis that we did, we are somewhere around $4. 50 per terabyte per month, which is almost half of the price as compared to other data platforms out there. So I think that's a great TCO to deliver right off the back. And as you scale from one exabyte to multiple exabytes, there's a lot more to realize.

Dave Vellante

>> And again, this is primarily for training, or is it also for inference? How do you see that?

>> So for the enterprise, there's two sets of customers.

Garima Kapoor

>> One is enterprise customers that are merging the data and AI workloads together. So they're building massive data lakes in which they're trying to bring the AI workloads as well on the same set of consolidated storage, if I may say, so that's the enterprise workloads. Then of course there are likes of the world that we all know about who are doing the training at massive scale. And for them, they are very specific in their environment of what they want to do, how they want to build it out. So there are two very distinct set of customers that we are talking about, and AI Labs is general purpose enough to address both of those workloads.

Dave Vellante

>> Okay. So then now you've got the ExaPOD, you've got the reference architecture, you've got your partners involved. Are you now bringing in, whether it's enterprises or labs, are offering them time in the lab so they can run the test? So how does that work? How do I get time in the lab? What am I doing when I ... Yeah.

>> Wow. For anyone who's interested for access to the lab,

Avi Shetty

>> just go to storageforai. com and you have all the instructions.

Savannah Peterson

>> I just love that URL Git. You don't always ... Yeah.

Avi Shetty

>> Thank you to our head of branding sitting right there. And, yes, we made it easy for our partners. We disclosed this to our partners leading up to the launch last month, and pretty much unanimous alignment. We're talking to our OEM engineers. Everyone is working in AI, but no one really has access, right? And their eyes opens up when we say, "Hey, we'll give it. We have exabytes of storage with Solidigm. " In this example, on Supermicro dense servers with Intel CPUs on it. So you have pretty much the entire ecosystem being represented with NVIDIA GPUs 800- gigabit network with FI. And essentially, a representation of an actual data center. So, yes, our customers contact us, we just give them hours of credit where they can log in remotely, access the environment. And we've set this business case in such a way that it's financially viable for Solidigm as well, where our partnership with FarmGPU. I want to give a shout-out to FarmGPU, our AI data center and our neocloud partner who hosts the data center for us. But when we are not using it or our partners are not using it, we actually rent it out on the RunPod neocloud platform, which helps us subsidize some of our OPEX. So it's making us invest more.

>> Smart. - Yeah, it's the neocloud environment.

Dave Vellante

>> And your customers

Avi Shetty

>> or prospects, they could self-serve, you said, right?

Savannah Peterson

>> Yes. - They come in through a VPC and-

Avi Shetty

>> It's just like access to any neocloud.

Dave Vellante

>> You log in to the data center,

Avi Shetty

>> the system has all the models available for you to download, test your environment, benchmark. And then once you're done, the pod clears itself up available the next one to use.

>> And you published these

Dave Vellante

>> results, or are they sort of private?

>> Yes. So we did an initial publication as part of MLPerf

Avi Shetty

>> where we were able to log in the highest performing single client throughput ever recorded at 120 megabytes per second on a single node. There were a bunch of other competitors on that list. The neighboring one, he was using 10 clients. The one next to us was using 15 clients to achieve similar scores, thus showcasing the advantage of optimizing the full stack, the storage, the software ISV, the OEM platform, as well as the CPU stack, thus getting appreciable MLPerf performance record with a single node. We are not stopping. We want to go work with MinIO to go record 1, 2, 4, 8, 16 node performances, which we are going to publish soon in a white paper or a POC setting.

>> How long have you all been working together?

Savannah Peterson

>> Tell me a little bit about the partnership. You can tell there's a lot of love there. And so I am curious, Garima, why don't you start with that?

>> I think we've been working together

Garima Kapoor

>> for quite some time now. And the thing is that because we are software-defined, so we need to leverage the underlying hardware in the best possible way. So for us, it's extremely important to partner with the likes of Solidigm, Intel, Supermicro to make sure that we are able to optimize and saturate whatever is given to us from hardware perspective. And that's where the beauty of software defined comes, that you can run it on any environment and be able to derive as much as performance as you can. And with Solidigm, I think one of the exciting things is in terms of the density of the drives and shrinking of data centers, because that's what we hear in terms of day in and out in our conversation with exascale customers, right? Because at that point, every single dollar counts. At that point, power consumption counts. At that point, data center footprint counts, everything. Every single dollar counts at that point. And if we can even bring it down by couple of percentage points, it goes long way. So, yeah, this partnership has been long in making and we are very, very happy and delighted to partner with Solidigm. And, yeah, the drives and the density and the roadmap of the density that we see coming along, I think you were sharing that it's 256 terabyte drives.

>> No, you're already outperforming yourself.

Savannah Peterson

>> Next year is- - It's awesome.

Savannah Peterson

>> You're high achievers on this. Cheers.

Garima Kapoor

>> Yes. - A shout-out to Intel. I think Intel came in.

Avi Shetty

>> We've been working with MinIO for the last as purely

Garima Kapoor

>> as a supplier, but more of a collaborative engineer to engineer discussion happened when Intel came along and said, "Hey, we want to build an ExaPOD, like an exabyte reference design. " And I think Intel got MinIO. We were still not in, we had to get away in with Supermicro as being the other platform provider. And I feel this was the right partnership across all the four ecosystem partners to demonstrate. And we worked closely for the last, I believe this happened after GTC this year.

>> Yes. . The one that-

Savannah Peterson

>> March?

Garima Kapoor

>> The March. Yeah. - Yeah, the March one.

Avi Shetty

>> I feel like we got a GTC

Dave Vellante

>> like every six days.

Garima Kapoor

>> I mean for good reason. That's no distant video, but-

Avi Shetty

>> And it led to the announcement, I think, as part of this conference from them, and we are continuing. We have a lot more benchmarking to do and, yeah, we'll keep generating more data.

Dave Vellante

>> You talked about sort of saturating the GPU in a good way. In your tests that you're running in the lab, how are you finding the testers, the enterprises, their ability to actually utilize the GPU? I believe that you've solved the storage problem. Are there other things that you're able to uncover that can inform best practices so organizations can better utilize these expensive resources? Have you found anything there, or is it predominantly the ability of the IO to saturate the GPU that is the bottleneck?

Avi Shetty

>> No, I think there are multiple bottlenecks across the networking. I'm not a network expert, but it's given me a ground reality very quickly. It's one of the most toughest components in this thing. And what we've also developed, Dave, is working with our partners from GPU, has built a custom telemetry module for us, which is tracking everything from GPU flops to storage, smart attributes and everything in between. And as a result, whenever the workload peaks, we have engineers speaking, "Hey, why did that happen? What's the firmware dependency to that? " We are trying to learn AI workloads from a storage- centric standpoint, the same telemetry. I think MinIO looks at it from a storage optimization standpoint and Intel looks at it from a CPU-centric standpoint. I think that's helped us uncover a few things and fix it before it shows up.

Dave Vellante

>> What kinds of things do you see your prospective customers and your customers doing? You hear a lot about coding, you hear a lot about contact centers.

Savannah Peterson

>> How are they making it real? I hear what you're saying.

Dave Vellante

>> Yeah, how do you make it real? I mean, in this show, you just see such amazing scientific discoveries. What are you seeing that is exciting you and your partners?

Garima Kapoor

>> Yeah, and just to close the last point that Avi was telling in terms of the GPU throughput, I think in any AI infrastructure, GPUs are the most expensive piece of it, and you want to keep them busy. So the throughput and the saturation becomes extremely important. And from story standpoint, it's our responsibility to push as much as data that we can to GPUs and networking becomes one of the bottlenecks in that perspective. That's why 800 gig network is so exciting for all of us. And then when you add it with S3 over RDMA, you're really able to saturate it because there you can see 2x difference know HTTP versus RDMA performance. That's where it really, really shines. So that's the optimizations from the story standpoint that we have done to saturate the GPU. Coming to the applied AI part, I think that's one of the things that we will continue to see innovation in next two to three years from enterprise perspective, how they start to consume. The exciting part, there is a couple of organizations that are doing applied AI right now. Autonomous vehicles are great. They're coming up with their robotaxi, Optimus robots. I think that's the real use case scenario where we see the lot of data growth and real applied AI that's going to impact all of us in a very, very meaningful way. Healthcare is another part where we see a lot of the applied AI getting used. So there are very specific verticals who are leveraging applied AI. But in terms of general purpose, I do feel that the training models that are happening, that's still going to be the predominant one for another one to two years till we see applications getting built out on top of them.

Savannah Peterson

>> You covered kind of the whole universe with that answer, which I really like. Avi, I mean, I'm curious, you mentioned you announced the lab in October, correct?

>> Yes. - Who has been most excited?

Savannah Peterson

>> What types of customers? What's the feedback been like?

Avi Shetty

>> Yeah. So we have our OEM customers who want

Avi Shetty

>> to understand the SSD solutions, which they deploy eventually. So they want a set of data sets. We have the end customers, the CSPs, the neoclouds who's like, "Hey, before I deploy, I want to get a sense of how the performance is going to look like for a 10 petabyte solution or a one petabyte solution with solid mSSDs. " So Solidigm management needs to be given credit for this because 120, 190, 220 drives is a big commitment from Solidigm for the ecosystem.

>> Absolutely. - And having that density available

Avi Shetty

>> to the customers for benchmarking POCs, pathfinding,

Savannah Peterson

>> collaborative work is what our customers partners, our immediate ecosystem community are very excited to come in. And in fact, we did another round of investments. I think you were asked about ... We currently have B200s and H100s in the lab. We are going to get liquid cooled B300s in the lab soon to go after the thermals, and the power, and the cooling collaboration activities.

Dave Vellante

>> You got to keep it refreshed. That's part of the commitment. It's not a one shot deal. People want the latest and greatest.

>> Yeah, the clock starts ticking when you deploy them. Yeah.

Dave Vellante

>> Absolutely. Yeah. - It's already behind.

Garima Kapoor

>> Yeah, exactly. - That's what it feels like something.

Savannah Peterson

>> I have one final question for you, Avi.

Savannah Peterson

>> I'm going to tailor this a little bit differently for you.

Garima Kapoor

>> So I'm going to let Garima go first. When we're hanging out at Supercomputing 2026. What do you hope to be able to say then that you can't yet say today?

>> That's a very interesting question.

Garima Kapoor

>> I've not really thought about it, but I think what we would be ... I'm hopeful discussing would be more on the applied AI side of things as applications become more mature in the AI space and what kind of read throughput, read requirements that they will have. I think what we are seeing right now is a mixed workloads of both read and write when it comes to checkpointing of data reads, when it comes from in terms of training of the data and so on, so forth. So I think we'll see a lot more evolution and lot more condensation of that workload overall. And then I think a lot more formalization around the applied AI part. I think that's still missing right now. So hopefully, in a year's timeframe, we'll be more talking about the application side and what can leverage the storage at exabyte scale. I think that that's where I'm hopeful that we'll have conversation.

>> I think you're spot on. I think you're absolutely spot on.

Savannah Peterson

>> Avi, since we talked about this earlier, I'm going to push you farther into the future. What are we going to be talking about in 5 to 10 years that we're not talking about right now? See, it gets harder. The problem is when we have you on repeat segments, I get more challenging with the closing question.

Avi Shetty

>> I have a standard answer for that one. I believe Jensen's quote where he talked about skills like people skills development and upskilling yourself where employees with AI, there will be jobs for employees with AI versus not, right? So there is always a question of, will AI take my job away? People with AI will take a job away. That's I think a fundamental change in how we are all looking at things where everyone from our engineers to our design engineers, to our supply chain people, our sales people are all upskilling themselves with using new tools to better their jobs. And I think that's going to continue. You're going to see a lot more software coming into the enterprise software stack, which betters our day-to-day job. And 10 years from now, I think, yeah, hopefully ... Waymo has been testing at my Palo Alto location for the last six months. We just want them to start offering.

Savannah Peterson

>> We're all ready for it. - Driverless

Avi Shetty

>> taxis, let's start with that.

Savannah Peterson

>> I feel you on the Peninsula taxis. I'm in Pacifica and they don't go past Daly City and it drives me nuts. Anyway, it's another thing. We could have a whole Waymo moment. You just said something. I'm sorry, I have a bonus question for you because you just said something that I think is interesting. So yesterday, I had the privilege of being on the University of Michigan campus in Ann Arbor talking to some students, some students studying this, which is so cool. Obviously, there's a lot of students at this event, so this is really relevant. And one of the questions that we had crafted to ask them is what they think is the most valuable human skill to have right now, because the job market is changing and a lot is happening. So I would love to ask you two, what do you think is the most valuable human skill? Avi, I'm going to start with you and then give Garima a little extra time since I put her on the spot for the last one.

Avi Shetty

>> I think the fundamentals don't change even with AI or without. You need to have strong, basic, fundamental understanding of whatever your job skill is when it comes to electrical engineering. You need to not know the underlying physics of how a transistor operates. AI is going to help you with all the things, but for you to be a better designer, you need to understand that, yes, white coding, peer programming will all help, but the fundamental logic doesn't change. That's you, which brings in. So I think that skill, everyone needs to be passionate and investing time in, but yeah, technology is there to assist, not to depend. So I think we keep going on that front.

Savannah Peterson

>> Yeah. Garima, what do you think?

>> Yeah. No, well said. I'm completely in agreement with

Garima Kapoor

>> what Avi just mentioned. I think the rules are going to get consolidated. IEC and leadership role are going to become one, and it's very important when those roles merge, you need to have the clarity of thought and clarity of thought can only come when you are expert in your domain. That's just the nature of it. You need to be able to understand end to end stack, as Avi was mentioning earlier, because otherwise you don't know what the end result is going to look like and you cannot ask AI to get the job done at the end of the day. So you need to know what the end result is going to be, and then AI can help you get there. So I think I tell my kids also they're in high school and they're like, "What do I study for future? " And I'm like, "Just read a lot of things and find what you're passionate about and be curious. " That's the only thing that I can guide you. There is nothing else right now that I can tell you go this way or another. Yeah.

Savannah Peterson

>> Be curious has always been a great. I was talking about the curiosity quotient. We got EQ/IQ and curiosity quotient, and I think the higher your curiosity quotient is right now.

Avi Shetty

>> Maybe same question to you both. You guys interview a lot of business leaders across different conferences. What's your take on that? The same question.

Dave Vellante

>> I think my answer would be the ability to read a room. I don't know if AI can read a room in a setting and can understand what the different perspectives are and how to draw them out and have that interaction. I don't think AI today anyway can do that.

Savannah Peterson

>> I think most kind of similar. My heart response when you said that is empathy. Because someone has going to have to tell the machines how to reason when it comes to-

>> Genuine empathy.

Savannah Peterson

>> Yeah, I like that. - Hey, that's a great question.

Savannah Peterson

>> Yeah. - Is it really?

Savannah Peterson

>> Wow. Yeah.

Dave Vellante

>> It's too late in the day for

Dave Vellante

>> me to go down to that path right now.

Dave Vellante

>> But I do think that empathy and compassion, because that's what's going to have to be that guides the machines to help solve the real problems we have in this world, not just do certain functions faster. We need to create better solutions, and that comes from the heart, not just the head. There's my closing sound bite for us. Yeah, that was a nice little moment. Thanks for asking us a question. Nobody ever asked me a question ever in this organization. I can count on one hand the number of times I've been asked a question at theCUBE. So that was very sweet of you. Thank you. Avi and Garima, fantastic way to close out the day. Thank you so much. And thank you, Dave. It's been a fantastic one here on the desk.

>> Yes, shout-out to Solidigm for making this day possible.

Savannah Peterson

>> Yeah, actually, just quick shout-out.

Dave Vellante

>> They're all hanging out over here. Solidigm team has done a really great job.

Dave Vellante

>> Straight to support. - You've curated a fantastic lineup

Savannah Peterson

>> of incredibly intelligent human beings that told a very dynamic story, and that makes our job fun and that also makes great content for our audience. So thank you all very much. Shout out obviously to the fantastic production team, because you were all smiling and looking cute there a second. You've been grinding. We've been on the road for seven weeks straight. You're doing great. I'm proud of you. And thank all of you for tuning in wherever you might be on this beautiful rock. We're here in St. Louis, Missouri at Supercomputing 2025. End of day one of three days of coverage. My name is Savannah Peterson. You're watching theCUBE, the leading source for enterprise tech news.