Nvidia B200 has a record in MLPERF while AMD MI325X competes with hopper plates

Nvidia and AMD released this week updated results from their AI accelerators at Benchmark Mlperf Inference. We have results for B200 Blackwell -based B200 and MI325X, the Instinct Architecture. Tests released include Llama and Stable Diffusion.

The tests are done with complete systems, each equipped with eight GPUs. In the case of Blackwell plates, for example, a NVIDIA DGX B200 was used, which reached the impressive score of 98,858 in the offline test and 98,443 online. It is 3 times more than the Hopper H200 generation scores and, comparing performance performance, it is 30 times more, according to Nvidia.

The AMD system with Mi325X GPUS, in turn, appears well behind. The timing of disclosure creates a tendency to compare with Blackwell, but in fact this generation of products aims to compete with Hopper, where performance is a little more pounded.

However, the instinct -based system cannot yet surpass the H200, even pulling 1000W per GPU, while Hopper uses 700W. Going to Stable Diffusion, AMD accelerators are behind even the last generation of Nvidia, the H100 Hopper.

AMD obviously already works in its next generation of AI accelerators, the Mi400X series, which promises considerable performance gains. However, even though the company is prepared to compete with Blackwell, Nvidia is already preparing for the arrival of its Blackwell Ultra -based B300 chips.

The gains between Blackwell and its Ultra version should not be as expressive as it has passed from Hopper, but represent an even greater challenge for AMD.

Related News:

It is important to remember, however, that leaving the front in performance is not the only way to compete. The cost x benefit is still an important metric for many companies and this is where AMD has sought to stand out more.

Fonte: WCCFTech