Hopper, Startups Debut MLPerf Scores

The MLPerf 2.1 inference release includes preliminary results that put Nvidia’s Hopper H100 in the performance lead. Asian startups Biren and Sapeon also made impressive debuts.

Linley Gwennap

The first public benchmarks for Nvidia’s new Hopper GPU put it atop the ranking for per-chip performance across all six MLPerf Inference benchmarks. But the H100 data fell short of the company’s expectations. Furthermore, Nvidia didn’t submit any H100 results in a DGX system; it has since disclosed that system availability has been delayed until 1Q23.

Also part of the MLPerf 2.1 benchmark release, startup Biren’s first official scores made its BR104 GPU the only other chip to outperform Nvidia’s A100, the erstwhile leader. Biren published only two of the six models but bested the A100 on both, even while falling behind the newer H100. More impressively, the Chinese company’s GPU ran at only 300W TDP, using less power than either Nvidia chip.

The other new vendor in this release is Sapeon—a spinoff from SK Telecom, one of Korea’s largest service providers. The startup posted results only for ResNet-50, but on that test, it nearly matched Nvidia’s power efficiency. The release also included updated scores for Qualcomm’s Cloud AI 100 accelerator as well as edge-AI chips such as Nvidia’s Orin.

The initial Hopper results are impressive and, as we expected, raise the standard for AI performance leadership to a high level. Without Hopper, Nvidia would’ve lost the industry lead to Biren on two benchmarks. Nvidia continues to offer the broadest set of scores, demonstrating the capability of its fully operational software stack.