NVIDIA BENCHMARK RECORDS 2.6 - SECOND DEEPSEEK - V3 WEIGHT REFIT

What happened: NVIDIA's internal tests found its Reshard method cut DeepSeek-V3 weight refit time on DGX H100 to 2.6 seconds from 14.9, trimming a recurring transfer step in reinforcement-learning workflows. Key numbers: 671-billion-parameter model. Why it matters: Reshard matches tensor layouts across GPUs and routes only needed shards instead of rebuilding a full tensor, which NVIDIA says reduces repeated traffic and spreads transfers among GPUs. An open-source implementation is hosted in NVIDIA's NCCL extensions repository.

NVIDIA's internal tests found its Reshard method cut DeepSeek-V3 weight refit time on DGX H100 to 2.6 seconds from 14.9, trimming a recurring transfer step in reinforcement-learning workflows.

Reshard matches tensor layouts across GPUs and routes only needed shards instead of rebuilding a full tensor, which NVIDIA says reduces repeated traffic and spreads transfers among GPUs. An open-source implementation is hosted in NVIDIA's NCCL extensions repository.