Event ·Multi-source

NVIDIA Blackwell Ultra 创纪录:预训练 DeepSeek-V3 671B 达 1648 TFLOPS/GPU

NVIDIA Blackwell Ultra 在预训练 671B 参数的 DeepSeek-V3 模型上达到每 GPU 1648 TFLOPS 的性能。这一数字约为上一代 GPU 性能的 3 倍。该成果通过 Megatron-Core、TorchTitan 和 JAX 等框架的协同设计与持续软件优化实现。Blackwell Ultra 在该基准测试中刷新了世界纪录。

Current read

NVIDIA 的 Blackwell Ultra 训练 DeepSeek-V3 671B 比上一代快 3 倍,每 GPU 跑到 1648 TFLOPS,优化太狠了。

9 sources83° AI heatLast updated 2026/7/21 15:03:04

Evidence

9 sources
Related source 2Aravind Srinivas
View source
Related source 3@koltregaskes
View source
Related source 4arXiv cs.LG
View source
Related source 5marktechpost
View source
Related source 8LMSYS Org (SGLang)
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.