NVIDIA Blackwell Ultra Sets World Record for DeepSeek-V3 Pre-Training: 1,648 TFLOPs per GPU

NVIDIA's Blackwell Ultra achieves 1,648 TFLOPs per GPU on DeepSeek-V3 671B pre-training, 3x the previous generation, thanks to co-design and continuous software optimization across Megatron-Core, TorchTitan, and JAX.

• Nvidia
• DeepSeek-V3
• 训练性能
• Blackwell Ultra