Raja Koduri and Satoshi Matsuoka Discuss the Future of HPC at SC21

By Tiffany Trader

November 29, 2021

HPCwire’s Managing Editor sits down with Intel’s Raja Koduri and Riken’s Satoshi Matsuoka in St. Louis for an off-the-cuff conversation about their SC21 experience, what comes after exascale and why they are collaborating.

Koduri, senior vice president and general manager of Intel’s accelerated computing systems and graphics (AXG) group, leads the team that is designing and implementing the GPU-accelerated Aurora exascale system, destined for Argonne National Laboratory next year. And Satoshi Matsuoka, director for the Riken Center for Computational Science, is one of the principal architects of Riken’s Arm-based Fugaku system, the world’s number-one ranked supercomputer.

What are these Intel and Riken leaders doing together at SC? Watch the interview – or read the transcript – to find out!

Transcript (lightly edited):

Tiffany Trader: Hello, I’m Tiffany Trader, managing editor of HPCwire. And here with me today are Satoshi Matsuoka, director for the Riken Center for Computational Science, and Raja Koduri, senior vice president and general manager of the accelerated computing system and graphics group, AXG – which you can see on his jacket – at Intel Corporation. And we’re all together at SC21 in St. Louis. We’re going to have some discussion about what we’re doing here at the show, and then what’s ahead in the future. Here we are on the eve of exascale, the cusp of exascale, so: how did we get here and where are we going into the future? And what are the next steps?

Satoshi-san, you are the innovator and the designer for the top system, the Fugaku supercomputer, which has been the number-one system now for four cycles of the Top500 list. So you’ve got four of those number-one certificates, but you’ve also got number one on HPCG, and I believe also HPL-AI with already some exascale results for that. So how is your show going here at SC21?

 Satoshi Matsuoka: Firstly, you know, it’s live and being live, you realize, you get lots more interactions than you would otherwise get on Zoom. Of course, the interaction we’re having here was very careful [with] COVID precautions. … It’s been quite an experience trying to demonstrate what Fugaku can do, but also getting a lot of input for future directions, talking to various institutions, researchers, being reunited with my research colleagues, as well as being able to meet the luminaries, like Raja, to really just discuss intensively about where we go from here.

Trader: And Raja, you and your group have been hard at work on the Ponte Vecchio GPU, which will be the centerpiece of the Aurora supercomputer, one of the first U.S. exascale systems coming online. And you’ve been here at the show – you and your colleagues have been talking about Ponte Vecchio, releasing just a few more details about that. And also talking about OneAPI. So how’s the show going for you?

Raja Koduri: First off, I’m super glad to be here live. I came to the show, frankly, because a lot of our partners, customers, and folks like Satoshi-san traveled all the way here. So I couldn’t let go of the opportunity to meet them all face to face and live. Many good interactions as Satoshi-san said, there’s nothing like you know, face to face interactions and being able to discuss ideas for the next generation and where things are going for the next five years, and so forth. And congratulations once again to Satoshi-san for the number-one supercomputer in the world.

I’m hoping next year, we will post some numbers, some U.S. exascale computers will go to the top of the list; we will see. And of course you know Ponte Vecchio has a big role to play in that, and we are hard at work. We have, as Intel, 2022 Sapphire Rapids, Sapphire Rapids with HBM, Ponte Vecchio – big products for us to ship and make available to our customers. So in many ways for us, this Supercomputing is a precursor for 2022; 2022 is the big year for us. So things are going well, like I said, we are hard at work. Lots of software work going on with OneAPI to bring it up on new architecture, tuning workloads, etc, etc. Working with our partners at Argonne Lab, enabling lots of HPC applications, you know, all the usual tough work, the steep hills to climb at the end to ship a product – that’s what we are doing today.

Trader: And Satoshi, what’s coming up for you at Riken with Fugaku and the Arm A64FX chip? What are your plans going into the future?

Matsuoka: Well firstly, Fugaku is in production now. It had been in pre-production last year around this time of year, but it went into production in March of 2021. And since then, the utilization has gone up steadily to the extent that now the machine is pretty full, which is quite amazing. It’s a 160,000-node machine with like 8 million cores. It’s not doing one big run, but across many applications, we are getting significant utilization. So there is demand for an exascale machine. When we started to embark on this journey toward exascale, we said, who’s going to use exascale capability? But now we have exemplar applications that utilize the entire machine as instances. But also we have a group of application users, a variety of applications that basically fill up the machine on a regular basis. And we’re, of course, we’re constantly trying to improve the software stack… well, the hardware is pretty much set, but maybe there are some minor improvements we can make, but most importantly, we are still involved in really improving the software and system software applications, algorithms, and branching out to broader areas of utility of supercomputing through non-traditional areas, businesses and industry usage, using what we call the Society 5.0 in Japan, really transforming the society for sustainability causes. And then, of course, we will be using Fugaku partly to investigate future architectures. You always build a machine, the next-generation machine, using the current-generation machine as a proxy, and as a basis. So we are very busy embarking on these, and we are growing and hiring, so if you’re interested, work with us.

Trader: There’s your offer. That’s quite a vision, and it’s really tied with helping society, which is the most important thing. So Raja, there were recently some changes with the HPC, the Super Compute, Supercomputing group at Intel, what is your vision for the Super Compute group going forward?

Aurora blade rendering with two Sapphire Rapids HBM CPUs and six Ponte Vecchio GPUs, shown at Intel Architecture Day (Aug. 2021)

Koduri: Yeah. You know, we at Intel, with Pat Gelsinger coming back, have kind of a newfound passion for supercomputing, and for us, supercomputing is not only the big national supercomputers in the big labs. We believe supercomputing capability needs to be available to every human on the planet. And for that, we envision a future where exascale computing is accessible to everybody, which means it needs to come down in power, right, you need to get exascale in kilowatts. And the cost needs to come down. And, you know, we are busy at work creating technologies that integrate the components that make up this exascale supercomputer in 2022 and 2023, you know, down to levels where every university should be able to afford an exascale computer in a few years from now. And… that’s not just kind of bringing exascale down, but it’ll also enable the next order of magnitude computing for the big computing centers. Kind of everybody shudders at the word zettascale, the next 1000X, but, you know, that’s how in the technology industry, we roll. That’s what kind of keeps us coming into work and creating technologies. So, it’s an exciting phase with Pat Gelsinger back and we are all over supercomputing.

Trader: Wonderful. And so zettascale is the next milestone. Satoshi, you’ve had a clear vision for what exascale means to you, are you thinking about what zettascale means to you?

Matsuoka: Well, when we met exascale, it wasn’t about reaching Linpack exaflops. It’s about reaching milestones of orders of magnitude performance increase over machines we had back in 2010, which for us was a K computer, of course there were similar machines in the U.S. and Europe and so forth. So, by achieving two orders, perhaps, three orders of magnitude speed-up, we may be able to really increase the application performance and portfolio, and as Raja said, we can even make it much more accessible, solving the world’s biggest problems. So this is really the application-first philosophy we have. And by that, what we aim to do is to really look at and revisit application performance, because as much as you might think the supercompute community’s all about performance, we actually measure performance and how we analyze performance is still actually not up there to my standard, even compared to some other engineering communities that are much more disciplined. So what we plan on doing is to really see what the status quo will be, involving many people and the partners at DOD, collaborators in the industry, Fujitsu, Intel, other companies like your rival companies, to basically set the standard and then say, well, what if we increase the application performance – what are the requirements to increase the application performance by 10X, 100X, and then what kind of architecture do we need, work on the algorithm changes we need to do, so really be application first – and to drive to the next-generation machine. This is our philosophy, and it will be an action over the next ten years.

Koduri: If I may add: the workload, understanding the current workloads about to run on all these big machines, and, you know, the Fugaku, and Satoshi-san’s computers are already enabling a bunch of interesting studies, which we are following, you know, his work, his analysis, and all. They are all going to guide us on how to architect the next-generation machines. And real applications – there are opportunities, there are definitely opportunities for orders of magnitude, whether we can get them one order, two orders, three orders, it all depends on what inventions we do and how much we understand. And as Satoshi-san said here, our understanding of workloads is still not very good, especially at large scale. Right. You know, we’re just beginning that understanding, simulating all of these workloads – and I’m a workload person, so I’m super excited to kind of dive into the details of those bottlenecks.

Trader: So I have one more question, but it’s a big one. So we have with us one of the leading developers and champions for the Arm system, the Arm supercomputing system, and then we have over here, Raja, who’s leading the design for the Intel GPU-accelerated supercomputer/exascale system. Why are we together?

Koduri: Yeah, you want to go first?

Matsuoka: Well firstly, you have to realize supercomputing penetrates the broader IT market; the software ecosystem becomes really important. And of course, on the other hand, in order to achieve the acceleration, we always want to have some sort of specialization and customization. Even A64FX, it’s a standard Arm processor, but has acceleration inside the chip as a new vector extension, just like AVX-512. Even Intel CPUs, you think they’re CPUs, but because they’re like, let’s say 386 versus, you know, whatever, whatever Lake or Rapids, it’s a completely different processor, not only architecturally, but the instruction set, the acceleration features are enormous, even though it’s a CPU. Then also GPUs have become really general-purpose, so it’s no longer valid to consider GPUs as being special-purpose, they’re really general-purpose. A little different type of computing model, you know, but people can program GPUs fairly easily; there’s a robust software stack. So it’s all about the software ecosystem. And there are some choices, like Arm, x86, some GPUs with various programming models. But at the end of the day, it’s really important to have a very stable software ecosystem, maybe combination thereof. And then, you know, build… think about what can we extend rather than have some brand-new customized acceleration, which may bring us 10 times speed-up but has no software. So that’s why I think GPUs and Arm can interplay very [nicely].

Koduri: Yeah, the thing I’ll add is that our collaborations already on things like OneAPI, OneDNN are yielding great results. For example, in OneDNN, you know, Satoshi-san’s team did work that figured out how to use vector math to do systolic… And, you know, we took that and we productized it. And it’s available for everybody, whether it’s AVX-512, or anybody who has a vector engine can get additional performance because of collaborations like that. And similarly, we were just talking about all the auto-vectorization work Intel does in our compilers, that we just put it into GCC and LLVM. And, you know, that could be taken advantage of by Fugaku, right. And so to hit that 1000X that I talked about, this type of collaboration, figuring out what they understood with the machine, and some of the technologies in our roadmap that are coming up, like, you know, silicon photonics integration, for example, is super exciting for me, some interesting caches, memory integration. And we would love to engage with folks like Satoshi-san on the building of his next computer. So, if he chooses an Arm, or x86 or RISC-V, or GPU or a combination or something else, you know, he should make the choice based on what’s right for the particular machine and the work.

Matsuoka: You know, we got number one on the CosmoFlow on the MLPerf HPC benchmark, being able to train this scientific AI code with 80,000 nodes on Fugaku. The underlying core of the training module is OneDNN, which was done by Intel, but the Arm version was developed in collaboration with Fujitsu. So the Arm version of OneDNN is our making, but was done in heavy collaboration with Intel. So Intel is not just x86, you know, narrow-focused company now, I think they’re very open to our collaboration in various ways.

Koduri: And that’s the new Intel under Pat’s leadership; “open” is our top architect, our highest architect. And he’s going to lean in on that up and down the stack.

Trader: With regards to this partnership and collaboration, do you have specific goals that you will be sharing more about in the future? Raja, do you want to take that?

Koduri: You know, definitely, right. As Satoshi-san said, these programs are sometimes 10-year programs. And there are many technologies that we can jointly look at what makes sense, and when the time is right, if we land on some common collaboration point, we’ll definitely be talking to the world and announcing it and all that stuff. But it’s more understanding the workloads, understanding what the goals are for next-generation systems that we want to do and we share our technology roadmaps with full flexibility years in advance before we share with the rest of the world.

Trader: Satoshi-san, Raja, thank you for talking with me today. I look forward to the next time. Take care.


To view more HPCwire exclusive SC21 video interviews, including our interview with Jack Dongarra, go here. Interviews with SiPearl and Preferred Networks will be out soon.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

ARM, Fujitsu Targeting Open-source Software for Power Efficiency in 2-nm Chip

July 19, 2024

Fujitsu and ARM are relying on open-source software to bring power efficiency to an air-cooled supercomputing chip that will ship in 2027. Monaka chip, which will be made using the 2-nanometer process, is based on the Read more…

SCALEing the CUDA Castle

July 18, 2024

In a previous article, HPCwire has reported on a way in which AMD can get across the CUDA moat that protects the Nvidia CUDA castle (at least for PyTorch AI projects.). Other tools have joined the CUDA castle siege. AMD Read more…

Quantum Watchers – Terrific Interview with Caltech’s John Preskill by CERN

July 17, 2024

In case you missed it, there's a fascinating interview with John Preskill, the prominent Caltech physicist and pioneering quantum computing researcher that was recently posted by CERN’s department of experimental physi Read more…

Aurora AI-Driven Atmosphere Model is 5,000x Faster Than Traditional Systems

July 16, 2024

While the onset of human-driven climate change brings with it many horrors, the increase in the frequency and strength of storms poses an enormous threat to communities across the globe. As climate change is warming ocea Read more…

Researchers Say Memory Bandwidth and NVLink Speeds in Hopper Not So Simple

July 15, 2024

Researchers measured the real-world bandwidth of Nvidia's Grace Hopper superchip, with the chip-to-chip interconnect results falling well short of theoretical claims. A paper published on July 10 by researchers in the U. Read more…

Belt-Tightening in Store for Most Federal FY25 Science Budets

July 15, 2024

If it’s summer, it’s federal budgeting time, not to mention an election year as well. There’s an excellent summary of the curent state of FY25 efforts reported in AIP’s policy FYI: Science Policy News. Belt-tight Read more…

SCALEing the CUDA Castle

July 18, 2024

In a previous article, HPCwire has reported on a way in which AMD can get across the CUDA moat that protects the Nvidia CUDA castle (at least for PyTorch AI pro Read more…

Aurora AI-Driven Atmosphere Model is 5,000x Faster Than Traditional Systems

July 16, 2024

While the onset of human-driven climate change brings with it many horrors, the increase in the frequency and strength of storms poses an enormous threat to com Read more…

Shutterstock 1886124835

Researchers Say Memory Bandwidth and NVLink Speeds in Hopper Not So Simple

July 15, 2024

Researchers measured the real-world bandwidth of Nvidia's Grace Hopper superchip, with the chip-to-chip interconnect results falling well short of theoretical c Read more…

Shutterstock 2203611339

NSF Issues Next Solicitation and More Detail on National Quantum Virtual Laboratory

July 10, 2024

After percolating for roughly a year, NSF has issued the next solicitation for the National Quantum Virtual Lab program — this one focused on design and imple Read more…

NCSA’s SEAS Team Keeps APACE of AlphaFold2

July 9, 2024

High-performance computing (HPC) can often be challenging for researchers to use because it requires expertise in working with large datasets, scaling the softw Read more…

Anders Jensen on Europe’s Plan for AI-optimized Supercomputers, Welcoming the UK, and More

July 8, 2024

The recent ISC24 conference in Hamburg showcased LUMI and other leadership-class supercomputers co-funded by the EuroHPC Joint Undertaking (JU), including three Read more…

Generative AI to Account for 1.5% of World’s Power Consumption by 2029

July 8, 2024

Generative AI will take on a larger chunk of the world's power consumption to keep up with the hefty hardware requirements to run applications. "AI chips repres Read more…

US Senators Propose $32 Billion in Annual AI Spending, but Critics Remain Unconvinced

July 5, 2024

Senate leader, Chuck Schumer, and three colleagues want the US government to spend at least $32 billion annually by 2026 for non-defense related AI systems.  T Read more…

Atos Outlines Plans to Get Acquired, and a Path Forward

May 21, 2024

Atos – via its subsidiary Eviden – is the second major supercomputer maker outside of HPE, while others have largely dropped out. The lack of integrators and Atos' financial turmoil have the HPC market worried. If Atos goes under, HPE will be the only major option for building large-scale systems. Read more…

Everyone Except Nvidia Forms Ultra Accelerator Link (UALink) Consortium

May 30, 2024

Consider the GPU. An island of SIMD greatness that makes light work of matrix math. Originally designed to rapidly paint dots on a computer monitor, it was then Read more…

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Shutterstock_1687123447

Nvidia Economics: Make $5-$7 for Every $1 Spent on GPUs

June 30, 2024

Nvidia is saying that companies could make $5 to $7 for every $1 invested in GPUs over a four-year period. Customers are investing billions in new Nvidia hardwa Read more…

Nvidia Shipped 3.76 Million Data-center GPUs in 2023, According to Study

June 10, 2024

Nvidia had an explosive 2023 in data-center GPU shipments, which totaled roughly 3.76 million units, according to a study conducted by semiconductor analyst fir Read more…

AMD Clears Up Messy GPU Roadmap, Upgrades Chips Annually

June 3, 2024

In the world of AI, there's a desperate search for an alternative to Nvidia's GPUs, and AMD is stepping up to the plate. AMD detailed its updated GPU roadmap, w Read more…

Some Reasons Why Aurora Didn’t Take First Place in the Top500 List

May 15, 2024

The makers of the Aurora supercomputer, which is housed at the Argonne National Laboratory, gave some reasons why the system didn't make the top spot on the Top Read more…

Intel’s Next-gen Falcon Shores Coming Out in Late 2025 

April 30, 2024

It's a long wait for customers hanging on for Intel's next-generation GPU, Falcon Shores, which will be released in late 2025.  "Then we have a rich, a very Read more…

Leading Solution Providers

Contributors

Google Announces Sixth-generation AI Chip, a TPU Called Trillium

May 17, 2024

On Tuesday May 14th, Google announced its sixth-generation TPU (tensor processing unit) called Trillium.  The chip, essentially a TPU v6, is the company's l Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

IonQ Plots Path to Commercial (Quantum) Advantage

July 2, 2024

IonQ, the trapped ion quantum computing specialist, delivered a progress report last week firming up 2024/25 product goals and reviewing its technology roadmap. Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

The NASA Black Hole Plunge

May 7, 2024

We have all thought about it. No one has done it, but now, thanks to HPC, we see what it looks like. Hold on to your feet because NASA has released videos of wh Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire