H100 MLPerf performance has improved by 54 percent since the September 2022 round as a result of software optimisations.

The company specifically called out its 31 percent improvement on the 3D-UNet medical imaging benchmark.

Nvidia L4 Tensor Core GPU results were submitted for the first time. This low-profile accelerator is designed for use in almost any server, and successfully ran all of the MLPerf workloads.

Relative to the prior-generation T4, the L4 delivered from 2.2 times the performance on the ResNet-50 benchmark to a better than threefold improvement on the BERT 99.9% benchmark.

Nvidia pointed out that the L4 also delivers up to ten times faster image decode, up to 3.2 times faster video processing and over four times faster graphics and real-time rendering performance.

Other results include improvements of up to 63 percent in energy efficiency and 81 percent in performance for the low-power Jetson AGX Orin module compared with its results from this time in 2022, while the Jetson Orin NX 16G delivered up to 3.2 times the performance of its Xavier NX predecessor.

Nvidia's MLPerf results are available here.