NVIDIA excels in MLPerf Training benchmarks with significant performance improvements.

From NVIDIA: 2024-06-12 11:00:30

NVIDIA’s accelerated computing platform excelled in the latest MLPerf Training benchmarks, with performance on the LLM benchmark more than tripling compared to last year’s record-setting submission. Using 11,616 NVIDIA H100 Tensor Core GPUs, the platform achieved this feat through larger scale and full-stack engineering. Business opportunities from this AI performance are significant.

The NVIDIA H200 Tensor GPU, with 141GB of HBM3 memory, showed up to 47% performance increase in AI training in its MLPerf Training debut. Software optimizations have also led to 27% faster submissions using a 512 H100 GPU configuration compared to last year, with nearly perfect scaling observed with increased GPU numbers.

Enterprises can benefit from NVIDIA’s platform for LLM fine-tuning and Stable Diffusion v2 training, with up to 80% training performance acceleration seen in Stable Diffusion v2. The platform showcases high efficiency, as evidenced by the 47% boost in single-node GNN training with the H200 GPU compared to the H100.

Ten NVIDIA partners, including ASUS, Dell Technologies, and Oracle, participated in the benchmarks, highlighting the ecosystem’s wide support. MLCommons’ benchmarking efforts play a crucial role in providing companies with valuable data for AI computing decisions. The NVIDIA Blackwell platform promises next-level AI performance on trillion-parameter generative AI models.



Read more at NVIDIA: MLPerf Training Results Showcase Unprecedented Performance, Elasticity