Teortaxes says NVIDIA B300 NVL72 with Dynamo holds a 4x efficiency-per-watt advantage over DeepSeek DSpark
The analysis suggests the performance gap may close under capital-expenditure parity.
Story Brief
The analysis suggests the performance gap may close under capital-expenditure parity.
Source Posts on X
Cluster sources
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexTECH
8K/GPU at 80 tps on B300 NVL72 with Dynamo DeepSeek does 1K with DSpark on whatever it is they've got (and they do give me 80 tps now). Assuming it's H800 or either of 950s, that's at a minimum 4x difference per watt But honestly, might be parity per $ in capex https://x.com/teortaxesTex/status/2074622647724396855/photo/1 https://twitter.com/SemiAnalysis_/status/2074599515256918065
Zephyr@zephyr_z9