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

By Agam Shah

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 power it brings to artificial intelligence. 

Nvidia’s DGX GB200—interchangeably called the NVL72 rack server—can be configured with up to 576 GPUs within an NVLink domain. Beyond that, DGX systems can scale to tens of thousands of other GB200 systems through SuperPOD configurations with InfiniBand networking for long-distance communications. 

HPCwire spoke to Charlie Boyle, the vice president and general manager of DGX Systems unit at Nvidia, to understand the design and underlying technology of the system. Here’s an edited transcript. 

HPCwire: What is the new DGX system? 

Boyle: In 2016, when we launched the original platform at GTC, we introduced something the world hadn’t seen before. It was the first time we had NVLink, GPUs, and SXM. We’ve gone through multiple generations of DGX systems, from Pascal to Volta and Hopper, and have obviously been very successful with customers and all forms of enterprises worldwide.  

We announced a new type of system, just like we announced a new type of system eight years ago. This new system is a rack-level computer, which we call the DGX GB200 system. It has 72 Blackwells and 36 Grace GPUs, all integrated into a single NVLink domain. 

HPCwire: Clearly, more horsepower — can you share more about the decision to land at that configuration? 

Boyle: The original DGX has been an NVLink domain in a single chassis since its inception. The DGX-2 had a 16-way variant for a while, but even after that, we went back to an 8-way domain. As AI models become increasingly complex, especially with the new type of model called mixture of experts, where multiple AI models work together to answer a single question or generate a single output, we saw a limiting factor in practice.  

In the MLPerf example, one of the limiting factors was that very large models spent about 60% of their time communicating with each other inside the actual model. We realized that if we could build a much larger NVLink domain, it would alleviate that communication problem because NVLink is much faster than even the fastest InfiniBand. 

With this new system, in a single rack, you get 72 GPUs, 36 CPUs, and nine NVSwitch units, all delivered to the customer as a single DGX unit. The new rack system is also liquid-cooled, making it extremely energy-efficient, and it can take ambient plus input water for datacenter savings.  

Nvidia’s DGX GB200 (NVL72) Front shows Blackwell compute and switch trays, rear view of bundled cables carrying the 130 TB/second of NVLink between nodes (Source: Nvidia)

HPCwire: How far can you scale beyond one GB200 system? 

Boyle: The single system, which we call the DGX GB200 system — you may also see it referred to as NVL-72 — can be connected with as many racks as you want, going up to tens of thousands of GPUs in a DGX SuperPOD configuration.  

The SuperPOD product is a turnkey offering from Nvidia. We build it, ship it to the customer, and install everything, including customer acceptance testing. 

The big new thing in this SuperPOD is that previous generation SuperPODs were assembled on-site at the customer’s location with the classic DGX systems you’re used to seeing.  

This new SuperPOD will be built completely in the factory, run through all the burn-in tests, have half the cables removed, and then the racks will be shipped to the customer. Once there, the cables will be plugged back in, and final customer acceptance will be done for an even faster time to value. 

HPCwire: For those who don’t know, what is the difference between NVLink and InfiniBand? 

Boyle: NVLink is an inter-chip communication technology that operates like a memory system. It functions like a memory bus in its semantics and how you perform operations.  

Due to its specialized nature, NVLink is much faster than InfiniBand but also has distance limitations because of speed and latency maximums. NVLink has features that allow it to quickly communicate with all the chips simultaneously. In contrast, traditional networking technologies like InfiniBand send information in a classic data source-destination manner. 

In the 72 GPU rack configuration, any GPU can directly access the memory of any other GPU as if it were local to it. NVLink and InfiniBand are complementary technologies. Previously, the maximum NVLink domain was on a physical board called the HGX board, which had eight GPUs and a couple of NVSwitch units, with all the NVLink connections done on PCB traces. 

In this new generation, we were able to expand the NVLink domain from the compute chips themselves, which are the Grace-Hopper Superchips.  

When you see a picture of the system and look at the back of it, you’ll notice the compute trays. Each compute tray has two Grace CPU modules and four Hopper modules. On the back of each tray are external NVLink connectors. You’ll see a fully cabled backplane. Every NVLink port on the back of the compute shelf is directly connected to an external NVSwitch in the same rack, with all the connections happening at the back of the system. 

On the front of the system, you’ll find all the normal networking connectivity. Each tray has four InfiniBand ports and includes our BlueField-3 technology for north-south communication. BlueField can run in either InfiniBand or Ethernet mode. The tray also has all the standard management and Ethernet ports. The rack itself includes the networking infrastructure for management within the rack. 

GB200 Compute and NVLink interconnect tray (Source: Nvidia)

HPCwire: Will the new system change the way CUDA programmers write code for Nvidia systems? 

Boyle: NVL72 is the more common configuration for the new system, where everything is designed to work together seamlessly. One of the core libraries that has been super important is NCCL (NVIDIA Collective Communications Library). Over the years, we’ve improved NCCL and enhanced how it understands the different elements in the system. 

In this new system, as a CUDA programmer or a higher-level programmer, you can access all the GPU memory from any application. Out of the box, you are getting a 4x training improvement on Hopper and above a 30x improvement on inference compared to the previous generation. The best part is that it runs all the same software as you had on the A100 without needing any special programming semantics to harness the power of the new system. 

One of the big things we’re doing with software is not exposing that complexity to people if they don’t want it. Of course, you can still program directly at a lower level if desired. However, you can also run a simple [PyTorch] command … and the system will automatically place all that work across all the OS images running in the rack, ensuring optimal placement to get the job done. 

We’ve removed a lot of that complexity from the software, considering that enterprises that generally buy lots of DGX systems are increasingly getting into this field. They have data scientists and people who want to run AI, but they don’t necessarily have individuals who want to program directly to the chip at a low level. Obviously, we have some customers who do that, and we fully enable that technology for them. 

HPCwire: How do you think of the DGX design going forward? What did you learn from past designs? 

Boyle: As we look at the GB200 system being the flagship going forward, a lot of what we are putting in the system are things that hopefully customers never have to see. We’ve learned a lot from building very large clusters because NVIDIA does that independently for researchers.  

We’re adding a lot of predictive maintenance, job automation, and telemetry into the system so that it takes care of itself. As the systems get increasingly more complex and the user base expands more and more into other aspects of the enterprise, many enterprises don’t have a datacenter to put these systems into, so they’ll be placed with one of our datacenter providers. We’ve got a whole DGX program for that. 

In the future, customers will be running extremely complex jobs, and any little hiccup in the entire cluster could take down the job. However, the customer just wants to get work done.  

One of the advances that we’re working on in this platform, which is a combination of hardware and software… there are specific new RAS (Reliability, Availability, and Serviceability) features in the chip to help us predict what’s going on.  

We have a predictive maintenance AI that we run at the cluster level to see which nodes are healthy and which aren’t. More than just a binary “this is healthy, this isn’t,” we’re looking at the trail of data from all those GPUs, monitoring thousands of data points every second to see how the job can get optimally done. 

From a system design perspective, our original DGX goal was to build something that couldn’t be built at the time. In 2016, an 8-way NVLink system was unheard of, but now it’s the standard, and every CSP is building them. However, we still build them, and when we look to the future, it’s about building larger and larger clusters while making the cluster smart enough to execute the job that the customer wants it to do, taking care of all the little things that always happen in the cluster. 

If you’re running large-scale systems, something is always happening. We want to build the smarts into the cluster itself so that it takes the prime directive if you will: Get Work Done. If the job dies, we want to minimize restart time. On a very large job that used to take minutes to potentially hours, we’re trying to get that down to seconds. 

HPCwire: Cloud providers have multi-way DGX configurations similar to your systems. Is that how it works? 

Boyle: That’s our goal with DGX. We build it as a design reference and use it internally, but we also share that information broadly with all our partners.  

[The] number of cloud providers picking up GB200 GPUs all start with that reference design because it saves them a lot of time and money in R&D. 

They look at the reference design and say, “This is great, but I need it a little bit taller, I need a different manifold, and I want to use my own system management.” That’s the idea behind it. We’ve published the GB200 architecture to all of our partners, and they’re all building systems based on it. 

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!

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 latest weapon in the AI battle with GPU maker Nvidia and clou Read more…

ISC 2024 Student Cluster Competition

May 16, 2024

The 2024 ISC 2024 competition welcomed 19 virtual (remote) and eight in-person teams. The in-person teams participated in the conference venue and, while the virtual teams competed using the Bridges-2 supercomputers at t Read more…

Grace Hopper Gets Busy with Science 

May 16, 2024

Nvidia’s new Grace Hopper Superchip (GH200) processor has landed in nine new worldwide systems. The GH200 is a recently announced chip from Nvidia that eliminates the PCI bus from the CPU/GPU communications pathway.  Read more…

Europe’s Race towards Quantum-HPC Integration and Quantum Advantage

May 16, 2024

What an interesting panel, Quantum Advantage — Where are We and What is Needed? While the panelists looked slightly weary — their’s was, after all, one of the last panels at ISC 2024 — the discussion was fascinat Read more…

The Future of AI in Science

May 15, 2024

AI is one of the most transformative and valuable scientific tools ever developed. By harnessing vast amounts of data and computational power, AI systems can uncover patterns, generate insights, and make predictions that 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 Top500 list of the fastest supercomputers in the world. At s Read more…

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…

Europe’s Race towards Quantum-HPC Integration and Quantum Advantage

May 16, 2024

What an interesting panel, Quantum Advantage — Where are We and What is Needed? While the panelists looked slightly weary — their’s was, after all, one of Read more…

The Future of AI in Science

May 15, 2024

AI is one of the most transformative and valuable scientific tools ever developed. By harnessing vast amounts of data and computational power, AI systems can un 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…

ISC 2024 Keynote: High-precision Computing Will Be a Foundation for AI Models

May 15, 2024

Some scientific computing applications cannot sacrifice accuracy and will always require high-precision computing. Therefore, conventional high-performance c Read more…

Shutterstock 493860193

Linux Foundation Announces the Launch of the High-Performance Software Foundation

May 14, 2024

The Linux Foundation, the nonprofit organization enabling mass innovation through open source, is excited to announce the launch of the High-Performance Softw Read more…

ISC 2024: Hyperion Research Predicts HPC Market Rebound after Flat 2023

May 13, 2024

First, the top line: the overall HPC market was flat in 2023 at roughly $37 billion, bogged down by supply chain issues and slowed acceptance of some larger sys Read more…

Top 500: Aurora Breaks into Exascale, but Can’t Get to the Frontier of HPC

May 13, 2024

The 63rd installment of the TOP500 list is available today in coordination with the kickoff of ISC 2024 in Hamburg, Germany. Once again, the Frontier system at Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. 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…

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…

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…

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... 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…

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…

Leading Solution Providers

Contributors

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it Read more…

Eyes on the Quantum Prize – D-Wave Says its Time is Now

January 30, 2024

Early quantum computing pioneer D-Wave again asserted – that at least for D-Wave – the commercial quantum era has begun. Speaking at its first in-person Ana Read more…

The GenAI Datacenter Squeeze Is Here

February 1, 2024

The immediate effect of the GenAI GPU Squeeze was to reduce availability, either direct purchase or cloud access, increase cost, and push demand through the roof. A secondary issue has been developing over the last several years. Even though your organization secured several racks... Read more…

Intel Plans Falcon Shores 2 GPU Supercomputing Chip for 2026  

August 8, 2023

Intel is planning to onboard a new version of the Falcon Shores chip in 2026, which is code-named Falcon Shores 2. The new product was announced by CEO Pat Gel 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…

GenAI Having Major Impact on Data Culture, Survey Says

February 21, 2024

While 2023 was the year of GenAI, the adoption rates for GenAI did not match expectations. Most organizations are continuing to invest in GenAI but are yet to Read more…

How the Chip Industry is Helping a Battery Company

May 8, 2024

Chip companies, once seen as engineering pure plays, are now at the center of geopolitical intrigue. Chip manufacturing firms, especially TSMC and Intel, have b 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…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire