NVIDIA says Blackwell Ultra hit 1,648 TFLOPs per GPU on DeepSeek-V3 671B training
NVIDIAAI · x · 2026-07-21
NVIDIA says Blackwell Ultra set a new throughput record on DeepSeek-V3 671B pre-training, reaching 1,648 TFLOPs per GPU.
- NVIDIA says this is about 3× the delivered throughput of the previous generation.
- The company attributes the gain to co-design and software optimization across Megatron-Core, TorchTitan, and JAX.
- The chart compares GB200 NVL72 with GB300 NVL72, showing the jump from 606 to 1,648 TFLOPs/GPU.
Related event: NVIDIA Blackwell Ultra Sets DeepSeek-V3 Training Throughput Record(2 posts)→
More from Infra
- CoreWeave 5-year CDS spikes to ~1,100bps as market fears AI bubble starts as credit event — kimmonismus · 2026-10-26
- NVIDIA Plans ~$2,000 Per GPU Per Year Software Fee for Running AI Agents — LinusEkenstam · 2026-10-26
- 12 DevOps concepts to learn in order, from Linux to SRE — _jaydeepkarale · 2026-10-26
- Killed our Vector DB for SQLite BM25 and wrote a 3.6ns Rust sentinel: lessons from multi-agent loops — OpenComplaint524 · 2026-10-26
- Gemma 4 26B-A4B runs offline on a 16 GB Mac: 33/40 on TypeScript tasks at 31 tok/s — Mission-Sir-1614 · 2026-10-26
- How thousands of GPUs train one AI: a breakdown of 5D parallelism — jbhuang0604 · 2026-10-26