The MLPerf benchmark results, which reflect the time it takes to train various AI models on processors, continue to reinforce the notion that Nvidia is not only ruling the AI chip arena, but continuing to grow stronger. The company led all nine workload categories of the MLPerf Version 4.0 Training benchmark submissions, continuing a streak of MLPerf dominance that stretches back to the 2018 creation of the benchmark.
“There are nine workloads all told as part of MLPerf, and we set new records on five of those nine workloads,” said Dave Salvator, director of accelerated computing products at Nvidia.
He added, “But in addition, we are constantly optimizing and fine-tuning our software. We are actually publishing our containerized software on a monthly cadence,” which. he explained, shows that Nvidia can continue to boost performance on existing hardware architectures even as it accelerates its hardware roadmap to releasing new hardware families on a yearly basis.
“Even as we move to a yearly cadence with new hardware, we can show there is a lot of life left in the Hopper architecture,” Salvator said.
Among its latest MLPerf achievements, Nvidia established new large language model training performance and scale records on industry benchmarks with 11,616 Hopper GPUs. It also laid claim to “tripling training performance on GPT3-175B benchmark in just one year with near-perfect scaling.” Additionally, to Salvator’s point, Nvidia said it was able to increase speeds on its Hopper submissions speeds by nearly 30% via software enhancements.
Still, there may not be enough other companies touting and explaining (and spinning) their own MLPerf results for the market to figure out how much MLPerf shows them about what they should buy and what kind of overall performance they can expect as they increase their work with AI. While Nvidia continues to hold large press briefings and post blogs about ts MLPerf achievements, Intel is the only other company that now regularly issues public statements about its MLPerf showings, and it does do primarily to to point out how its Gaudi accelerators are a lower-cost but highly-effective alternative to Nvidia products.
This time around, Intel submitted benchmark results for its Gaudi 2 (It has not yet submitted anything regarding the recently announced Gaudi 3, which is thought to be a more direct Nvidia challenger).
Intel’s statement on MLPerf Version 4.0 read in part, “For the first time on the MLPerf benchmark, Intel submitted results on a large Gaudi 2 system (1,024 Gaudi 2 accelerators) trained on the Intel Tiber Developer Cloud to demonstrate Gaudi 2 performance and scalability and Intel’s cloud capacity for training MLPerf’s GPT-3 175B1 parameter benchmark model.
Zane Ball, Intel corporate vice president and general manager, DCAI Product Management, added, “The industry has a clear need: Address the gaps in today’s generative AI enterprise offerings with high-performance, high-efficiency compute options. The latest MLPerf results published by MLCommons illustrate the unique value Intel Gaudi brings to market as enterprises and customers seek more cost-efficient, scalable systems with standard networking and open software, making GenAI more accessible to more customers.”
Google also continues to submit and post its MLPerf results, but those are of primary interest to Google itself and some of its users, and not for the broader market.
Karl Freund, Founder and Principal Analyst, Cambrian-AI Research LLC, frequently provides more in-depth analysis of MLPerf results, and he did so this week at Forbes.