MLPerf has set itself the goal of establishing a better comparability for the determination and comparison of computing power in the AI ​​or ML area. In addition to the major chip manufacturers Intel and NVIDIA, ARM, Google, Intel, MediaTek, Microsoft and many other companies are also involved and thus enable better comparability of performance in this area.

After The results for MLPerf Training v0.7 were published some time ago, in which NVIDIA’s new A 100 took the lead, the results for MLPerf Inference v0 .7 online. This does not consider training with high accuracy, but rather the performance for inferencing, in which lower accuracy and high throughput are required.

The comparison systems are therefore limited to those with NVIDIA’s T4 and A 100 are equipped, as well as those with Xilinx FPGA and Intel’s Cooper Lake Xeons have made a contribution.

If you compare the results with each other, you have to make sure that systems with two, four, eight and more GPU accelerators are compared with systems with only one FPGA or two, four or eight processors. So you should keep a certain scaling factor in mind. In addition, the results are not the same across all applications. Depending on the software and interfaces, sometimes one platform benefits and sometimes the other.

However, it can be said that an A 100 -GPU accelerator can be more than 200 times faster than a Xeon processor based on Cooper Lake. Even the smaller T4, which is explicitly designed for inferencing, is sometimes faster by a factor of 30. NVIDIA is with its AI solutions (A 100, T4 and also Jetson AGX Xavier) in 85% of all benchmarks ahead.

Of course NVIDIA uses the MLPerf-P