For years, Nvidia has dominated many machine learning benchmarks, and now there are two more notches in its belt.

MLPerf, the AI benchmarking suite sometimes called “the Olympics of machine learning,” has released a new set of training tests to help make more and better apples-to-apples comparisons between competing computer systems. One of MLPerf’s new tests concerns fine-tuning of large language models, a process that takes an existing trained model and trains it a bit more with specialized knowledge to make it fit for a particular purpose. The other is for graph neural networks, a type of machine learning behind some literature databases, fraud detection in financial systems, and social networks.

Even with the additions and the participation of computers using Google’s and Intel’s AI accelerators, systems powered by Nvidia’s Hopper architecture dominated the results once again. One system that included 11,616 Nvidia H100 GPUs—the largest collection yet—topped each of the nine benchmarks, setting records in five of them (including the two new benchmarks).

“If you just throw hardware at the problem, it’s not a given that you’re going to improve.” —Dave Salvator, Nvidia

The 11,616-H100 system is “the biggest we’ve ever done,” says Dave Salvator, director of accelerated computing products at Nvidia. It smashed through the GPT-3 training trial in less than 3.5 minutes. A 512-GPU system, for comparison, took about 51 minutes. (Note that the GPT-3 task is not a full training, which could take weeks and cost millions of dollars. Instead, the computers train on a representative portion of the data, at an agreed-upon point well before completion.)

Compared to Nvidia’s largest entrant on GPT-3 last year, a 3,584 H100 computer, the 3.5-minute result represents a 3.2-fold improvement. You might expect that just from the difference in the size of these systems, but in AI computing that isn’t always the case, explains Salvator. “If you just throw hardware at the problem, it’s not a given that you’re going to improve,” he says.

“We are getting essentially linear scaling,” says Salvator. By that he means that twice as many GPUs lead to a halved training time. “[That] represents a great achievement from our engineering teams,” he adds.

Competitors are also getting closer to linear scaling. This round Intel deployed a system using 1,024 GPUs that performed the GPT-3 task in 67 minutes versus a computer one-fourth the size that took 224 minutes six months ago. Google’s largest GPT-3 entry used 12-times the number of TPU v5p accelerators as its smallest entry and performed its task nine times as fast.

Linear scaling is going to be particularly important for upcoming “AI factories” housing 100,000 GPUs or more, Salvator says. He says to expect one such data center to come online this year, and another, using Nvidia’s next architecture, Blackwell, to startup in 2025.

Nvidia’s streak continues

Nvidia continued to boost training times despite using the same architecture, Hopper, as it did in last year’s training results. That’s all down to software improvements, says Salvator. “Typically, we’ll get a 2-2.5x [boost] from software after a new architecture is released,” he says.

For GPT-3 training, Nvidia logged a 27 percent improvement from the June 2023 MLPerf benchmarks. Salvator says there were several software changes behind the boost. For example, Nvidia engineers tuned up Hopper’s use of less accurate, 8-bit floating point operations by trimming unnecessary conversions between 8-bit and 16-bit numbers and better targeting of which layers of a neural network could use the lower precision number format. They also found a more intelligent way to adjust the power budget of each chip’s compute engines, and sped communication among GPUs in a way that Salvator likened to “buttering your toast while it’s still in the toaster.”

Additionally, the company implemented a scheme called flash attention. Invented in the Stanford University laboratory of Samba Nova founder Chris Re, flash attention is an algorithm that speeds transformer networks by minimizing writes to memory. When it first showed up in MLPerf benchmarks, flash attention shaved as much as 10 percent from training times. (Intel, too, used a version of flash attention but not for GPT-3. It instead used the algorithm for one of the new benchmarks, fine-tuning.)

Using other software and network tricks, Nvidia delivered an 80 percent speedup in the text-to-image test, Stable Diffusion, versus its submission in November 2023.

New benchmarks

MLPerf adds new benchmarks and upgrades old ones to stay relevant to what’s happening in the AI industry. This year saw the addition of fine-tuning and graph neural networks.

Fine tuning takes an already trained LLM and specializes it for use in a particular field. Nvidia, for example took a trained 43-billion-parameter model and trained it on the GPU-maker’s design files and documentation to create ChipNeMo, an AI intended to boost the productivity of its chip designers. At the time, the company’s chief technology officer Bill Dally said that training an LLM was like giving it a liberal arts education, and fine tuning was like sending it to graduate school.

The MLPerf benchmark takes a pretrained Llama-2-70B model and asks the system to fine tune it using a dataset of government documents with the goal of generating more accurate document summaries.

There are several ways to do fine-tuning. MLPerf chose one called low-rank adaptation (LoRA). The method winds up training only a small portion of the LLM’s parameters leading to a 3-fold lower burden on hardware and reduced use of memory and storage versus other methods, according to the organization.

The other new benchmark involved a graph neural network (GNN). These are for problems that can be represented by a very large set of interconnected nodes, such as a social network or a recommender system. Compared to other AI tasks, GNNs require a lot of communication between nodes in a computer.

The benchmark trained a GNN on a database that shows relationships about academic authors, papers, and institutes—a graph with 547 million nodes and 5.8 billion edges. The neural network was then trained to predict the right label for each node in the graph.

Future fights

Training rounds in 2025 may see head-to-head contests comparing new accelerators from AMD, Intel, and Nvidia. AMD’s MI300 series was launched about six months ago, and a memory-boosted upgrade the MI325x is planned for the end of 2024, with the next generation MI350 slated for 2025. Intel says its Gaudi 3, generally available to computer makers later this year, will appear in MLPerf’s upcoming inferencing benchmarks. Intel executives have said the new chip has the capacity to beat H100 at training LLMs. But the victory may be short-lived, as Nvidia has unveiled a new architecture, Blackwell, which is planned for late this year.

Over just the past couple of years, supercomputing has accelerated into the exascale era—with the world’s most massive machines capable of performing over a billion billion operations per second. But unless big efficiency improvements can intervene along its exponential growth curve, computing is also anticipated to require increasingly impractical and unsustainable amounts of energy—even, according to one widely cited study, by 2040 demanding more energy than the world’s total present-day output.

Fortunately, the high-performance computing community is shifting focus now toward not just increased performance (measured in raw petaflops or exaflops) but also higher efficiency, boosting the number of operations per watt.

The Green500 list saw newcomers enter into the top three spots, suggesting that some of the world’s newest high-performance systems may be chasing efficiency at least as much as sheer power.

The newest ranking of the Top500 supercomputers (a list of the world’s most powerful machines) and its cousin the Green500 (ranking instead the world’s highest-efficiency machines) came out last week. The leading 10 of the Top 500 largest supercomputers remains mostly unchanged, headed up by Oak Ridge National Laboratory’s Frontier exascale computer. There was only one new addition in the top 10, at No. 6: Swiss National Supercomputing Center’s Alps system. Meanwhile, Argonne National Laboratory’s Aurora doubled its size, but kept its second-tier ranking.

On the other hand, The Green500 list saw newcomers enter into the top three spots, suggesting that some of the world’s newest high-performance systems may be chasing efficiency at least as much as sheer power.

Heading up the new Green500 list was JEDI, Jülich Supercomputing Center’s prototype system for its impending JUPITER exascale computer. The No. 2 and No. 3 spots went to the University of Bristol’s Isambard AI, also the first phase of a larger planned system, and the Helios supercomputer from the Polish organization Cyfronet. In fourth place is the previous list’s leader, the Simons Foundation’s Henri.

A Hopper Runs Through It

The top three systems on the Green500 list have one thing in common—they are all built with Nvidia’s Grace Hopper superchips, a combination of the Hopper (H100) GPU and the Grace CPU. There are two main reasons why the Grace Hopper architecture is so efficient, says Dion Harris, director of accelerated data center go-to-market strategy at Nvidia. The first is the Grace CPU, which benefits from the ARM instruction set architecture’s superior power performance. Plus, he says, it incorporates a memory structure, called LPDDR5X, that’s commonly found in cellphones and is optimized for energy efficiency.

Nvidia’s GH200 Grace Hopper superchip, here deployed in Jülich’s JEDI machine, now powers the world’s top three most efficient HPC systems. Jülich Supercomputing Center

The second advantage of the Grace Hopper, Harris says, is a newly developed interconnect between the Hopper GPU and the Grace CPU. The connection takes advantage of the CPU and GPU’s proximity to each other on one board, and achieves a bandwidth of 900 gigabits per second, about 7 times as fast as the latest PCIe gen5 interconnects. This allows the GPU to access the CPU’s memory quickly, which is particularly important for highly parallel applications such as AI training or graph neural networks, Harris says.

All three top systems use Grace Hoppers, but Jülich’s JEDI still leads the pack by a noticeable margin—72.7 gigaflops per watt, as opposed to 68.8 gigaflops per watt for the runner-up (and 65.4 gigaflops per watt for the previous champion). The JEDI team attributes their added success to the way they’ve connected their chips together. Their interconnect fabric was also from Nvidia—Quantum-2 InfiniBand—rather than the HPE Slingshot used by the other two top systems.

The JEDI team also cites specific optimizations they did to accommodate the Green500 benchmark. In addition to using all the latest Nvidia gear, JEDI cuts energy costs with its cooling system. Instead of using air or chilled water, JEDI circulates hot water throughout its compute nodes to take care of the excess heat. “Under normal weather conditions, the excess heat can be taken care of by free cooling units without the need of additional cold-water cooling,” says Benedikt von St. Vieth, head of the division for high-performance computing at Jülich.

JUPITER will use the same architecture as its prototype, JEDI, and von St. Vieth says he aims for it to maintain much of the prototype’s energy efficiency—although with increased scale, he adds, more energy may be lost to interconnecting fabric.

Of course, most crucial is the performance of these systems on real scientific tasks, not just on the Green500 benchmark. “It was really exciting to see these systems come online,” Nvidia’s Harris says, “But more importantly, I think we’re really excited to see the science come out of these systems, because I think [the energy efficiency] will have more impact on the applications even than on the benchmark.”