NVIDIA Blackwell achieves four times the performance of the H100 Hopper in MLPerf V4.1

The next graphic architecture NVIDIA Blackwell It hasn’t hit the market yet, but NVIDIA has already wanted to get the Artificial Intelligence sector excited showing the results that this platform gets in the new MLPerf v4.1.

Specifically, using one of the heaviest and most capable LLMs (large language models) “LLaMa 70B”, NVIDIA Hopper architecture delivers four times the performance of the NVIDIA H100.

To achieve this, this new architecture makes use of its second-generation Transformer Engine together with the FP4 Tensor Cores.

The Blackwell platform can be implemented in systems with up to 72 GPUs linked using NVLink technology thanks to the company’s switches.

The results also highlight: 27% performance increase in AI inference with the NVIDIA H200 GPU with the current “Hopper” architecture, a considerable improvement if we take into account that the hardware is the same and that everything is due to the evolution of the platform and the drivers and support from NVIDIA.

It thus joins AMD and Intel in showing results of its hardware.

End of Article. Tell us something in the Comments!