Preeti
NVIDIA’s full-stack accelerated computing platform has achieved exceptional performance in the latest MLPerf Training v4.0 benchmarks, more than tripling the performance on the large language model (LLM) benchmark compared to the previous year. This was accomplished using an AI supercomputer featuring 11,616 NVIDIA H100 Tensor Core GPUs connected with NVIDIA Quantum-2 InfiniBand networking. The scalability of the NVIDIA AI platform allows for faster training of massive AI models like GPT-3 175B, leading to significant business opportunities.
The NVIDIA H200 Tensor GPU, with 141GB of HBM3 memory and over 40% more memory bandwidth compared to the H100 GPU, has extended performance in AI training by up to 47% in its MLPerf Training debut. Software optimizations have also resulted in up to 27% faster performance with a 512 H100 GPU configuration compared to the previous year.
NVIDIA excelled at LLM fine-tuning and accelerated Stable Diffusion v2 training performance by up to 80%. The platform also showcased powerful performance in graph neural network (GNN) training, with the H200 delivering a 47% boost compared to the H100.
The broad ecosystem support for NVIDIA’s AI platform, with 10 partners submitting results, highlights the widespread adoption and trust in NVIDIA’s technology. MLCommons’ benchmarking efforts are crucial for enabling peer-reviewed comparisons of AI and HPC platforms, providing valuable data for purchasing decisions.
Overall, NVIDIA’s continuous software and hardware improvements have led to impressive performance gains in AI training and inference, with the upcoming NVIDIA Blackwell platform promising next-level AI performance on trillion-parameter generative AI models.
Source link
Source link: https://blogs.nvidia.com/blog/mlperf-training-benchmarks/
GIPHY App Key not set. Please check settings