IT之家 · 智能时代T2·78 pts
- 8,192 Ironwood chips pretrained DeepSeek-V3 671B in about 102 seconds, with 96% scaling efficiency at 4,096 chips
- Nvidia's 10,240-GPU GB300 cluster took the fastest time at 82.8 seconds, trading per-chip efficiency against cluster scale
- MLPerf debuted an agentic RL benchmark, where Ironwood ran Qwen3.5-397B to 76.5% coding accuracy