MLCommons, an industry group dedicated to artificial intelligence performance evaluation and machine learning hardware, Added result It added the latest artificial AI and ML accelerators to its database and essentially published the first performance numbers for Nvidia’s H100 and Biren’s BR104 computing GPUs through a series of industry-standard tests. The results were compared to those obtained with Intel’s Sapphire Rapids, Qualcomm’s AI 100, and Sapeon’s X220.
MLCommons’ MLPerf helps organizations deliver hardware test results MLPerf databaseThe MLPerf Inference version 2.1 set of benchmarks includes data center and edge usage scenarios, as well as image classification (ResNet 50 v1.5), natural language processor (BERT Large), speech recognition (RNN-T), medical imaging (3D) and other workloads. U-Net), Object Detection (RetinaNet), Recommended (DLRM).
Machines participating in these tests are evaluated in two modes. Queries arrive in bursts in server mode, whereas offline mode feeds all the data at once, so offline mode is obviously a better performer. Alternatively, the vendor can submit the results he obtained under two conditions. The closed category requires everyone to run mathematically equivalent neural networks, while the open category allows bids to modify them to optimize for hardware. IEEE spectrum.
The results obtained with MLPerf show not only the pure performance of the accelerators (e.g. 1 H100, 1 A100, 1 Biren BR104, etc.), but also their scalability, and per watt to draw a more detailed picture. It also describes the performance of While all results can be viewed in our database, Nvidia has compiled performance results per accelerator based on submissions from their own and third parties.
Some results are missing from the charts published by Nvidia, as Nvidia’s competitors have yet to submit all their results. Still, we can make a very interesting discovery in the table published by Nvidia (but keep in mind that Nvidia is a stakeholder here, so everything has to be taken into account).
Nvidia’s H100 is up to 4.5x faster than Nvidia’s A100 as it is the most complex and cutting-edge AI/ML accelerator underpinned by highly sophisticated software optimized for Nvidia’s CUDA architecture So it’s no surprise that it’s the fastest computing GPU out there today.
Still, Biren Technology’s BR104, which offers about half the performance set offered by the flagship BR100, shows considerable promise in image classification (ResNet-50) and natural language processing (BERT-Large) workloads. In fact, if the BR100 is twice as fast as his BR104, it will offer better performance than Nvidia’s H100 on image classification workloads, at least as far as performance per accelerator is concerned.
Sapeon’s X220-Enterprise and Qualcomm’s Cloud AI 100 can’t even compete with Nvidia’s A100, which launched about two years ago. Intel’s 4th Gen Xeon Scalable ‘Sapphire Rapids’ processors can run AI/ML workloads, but the code doesn’t seem to be optimized well enough for this CPU, resulting in pretty poor results.
Nvidia fully expects the H100 to offer even higher performance in AI/ML workloads, widening the gap to the A100 as engineers learn how to take advantage of the new architecture.
It remains to be seen how much computing accelerators such as Biren’s BR100/BR104, Sapeon’s X220-Enterprise, and Qualcomm’s Cloud AI 100 will significantly improve performance over time. Additionally, Nvidia’s H100’s true competitor is Intel’s codenamed Ponte Vecchio computing GPU, positioned for both supercomputing and AI/ML applications. It would also be interesting to see his MLPerf results for AMD’s Instinct MI250, which is primarily optimized for supercomputers. Yet Nvidia retains his AI/ML performance crown, at least for now.