NVIDIA noticed distorting the results of GPU RAPIDS tests and substituting the software algorithm – TH

In 2020, as part of the State of RAPIDS: Bridging the GPU Data Science Ecosystem, NVIDIA representatives talked about the benefits of their own set of GPU software libraries for open source software RAPIDS. The comparison was carried out with competing platforms, within which the green assured that the hardware system and software can cope with test loads 20 times faster than all solutions known on the market.

As it turned out, the results were obtained with violations. This fact was highlighted by members of the Transaction Processing Performance Council (TPC). It is an industry consortium that develops and creates benchmarks and benchmarks for servers. Its members are AMD, Intel and IBM. The essence of the complaint boils down to the fact that no one knows which test NVIDIA used during the presentation.

So, in such situations, the software developed by TPC is used, while NVIDIA used tests called “similar” to the official ones. Allegedly, an informal set of workloads was used, which may not correspond to the generally accepted ones. Such an approach could put NVIDIA in a privileged position, and the results are not valid and do not meet the standards.

At the moment NVIDIA is trying to solve the problem, but there are no more details on this matter.