MLPerf is growing at a rapid rate. Different market analysts, including Hyperion Research, IDC, and so on, all agree that this growth validates that MLPerf is indeed a helpful benchmark to make fair comparisons and purchasing decisions. These comparisons are especially useful to recognize that different vendors and OEMs collaborate to produce a high-performance system. Owing to its growth, the number of results that are submitted to MLPerf has grown approximately 30 percent for this round of submissions.
Furthermore, some customers use MLPerf as an entry point to assess the performance that they can expect from a new system that they plan to acquire. The acquisition might be a new Dell server or a new NVIDIA accelerator, or an overall system including the server and the accelerator. All these data points enable better decision making for data center design and hardware investments.
All Dell submissions to MLPerf Training v3.0 include NVIDIA accelerators. These accelerators enable a full spectrum of use cases that are seen in the MLPerf training benchmarks to run at faster time to convergence. With the introduction of new models such as DLRMv2, large language model LLM training has been an interesting driver for customers to reference. These large benchmarks model the data access patterns, compute consumption behavior, and other complex intricacies involved in training these workloads.
The following figure shows the percentage of submissions that are derived from the number of Closed division results. Dell Technologies submitted extensive results; we submitted a third of all results, followed by NVIDIA.
It is encouraging to see a substantial number of submissions from new submitters. This addition fosters more collective results and opportunities to obtain a specific result, if needed. For example, with a larger number of results, it is possible to obtain a specific datapoint about ResNet with a specific accelerator.