𝗡𝗩𝗜𝗗𝗜𝗔 𝗕𝗟𝗔𝗖𝗞𝗪𝗘𝗟𝗟 𝗗𝗢𝗠𝗜𝗡𝗔𝗧𝗘𝗦 𝗠𝗟𝗣𝗘𝗥𝗙 𝗧𝗥𝗔𝗜𝗡𝗜𝗡𝗚 𝟲.𝟬

NVIDIA won all seven benchmarks in the latest MLPerf Training 6.0 suite. The Blackwell platform achieved the fastest training times across every category.

The GB300 NVL72 system shows significant progress. It delivers 1.6x faster training than the GB200 NVL72.

Key performance data:

• GB300 NVL72 achieved 1.6x speedup over GB200 NVL72. • NVIDIA used NVFP4 precision to increase compute density. • The DeepSeek-V3 671B model trained on 8,192 GPUs via NVLink. • New Mixture-of-Experts workloads include DeepSeek-V3 671B and GPT-OSS-20B.

Large scale training requires massive communication between GPUs. NVIDIA uses fifth-generation NVLink Switches to connect 72 GPUs in one rack. This setup allows them to work as a single large GPU.

The company provides two networking options for these clusters:

NVIDIA remains the only vendor to submit results for all seven benchmarks. This performance helps them maintain their lead against competitors like Google and AMD.

The rapid improvement from GB200 to GB300 shows a fast engineering cycle. This speed sets a high bar for the AI training industry.

Source: blogs.nvidia.com

Optional learning community: https://t.me/GyaanSetuAi