NVIDIA Blackwell Ultra Sets DeepSeek-V3 Training Throughput Record

NVIDIA announced that its Blackwell Ultra GPU achieved a record-breaking 1,648 TFLOPs per GPU during the pre-training of the 671B-parameter DeepSeek-V3 model, highlighting significant advancements in large-scale AI training efficiency.

2026-07-21 ~ 2026-07-21 · 2 related posts

• NVIDIA says Blackwell Ultra hit 1,648 TFLOPs per GPU on DeepSeek-V3 671B training — NVIDIAAI · 2026-07-21

1 near-duplicate retellings: NVIDIAAI