DeepSeek V4.1F 在 Ascend 950DT 与 GB200 上的推理吞吐对比
DeepSeek V4.1F 跑在华为 950DT 上比 GB200 慢 5 到 7 倍，这篇拆解算了芯片单价和优化空间，推理党值得看看这笔账怎么算的。
根据 9 月 30 日基准测试，DeepSeek V4.1F+GB200 解码上限在 100 tps 下达 131K 上下文，而 950DT 配 CANN 仅 5K，差距 5 倍到 7.4 倍（230 tps 时为 20K 对 2.7K）。作者指出 DeepInfer 在 Nvidia 侧用了 EP64-128/DP64-128 大规模专家并行，而 DeepSeek 三周前在 DeepEP-Ascend 中承认 EP64-128 尚未优化，正与华为改进。若该优化在内部已解决，950DT 在 128 芯片超节点上可能提升到约 8-10K。按每 NVL72 机架 1000 万美元、每颗 950DT 传 3.7 万美元估算，盈亏平衡点约为每芯片 6.8K tokens/s，作者建议 DeepSeek 用不到 10 亿美元的小规模 Blackwell 集群覆盖 API，后训练推理全面转向 Ascend。
some thoughts. the current decode max for v4.1F+GB200 is 25.5K at 100 tps and 131K ctxt. 950DT with CANN, as per the Sep 30 benchmark, does 5K. At 230 tps: 20K vs 2.7K. From 5x to 7.4x. But: DeepInfer Nvidia uses EP64-128/DP64-128. 3 weeks ago, DeepSeek explicitly said in DeepEP-Ascend that they have not optimized EP64-128 (see pic 4, "Larger EP sizes and combine remain under optimization; combine has additional local reduction overhead and HBM contention with URMA. Huawei firmware improvements are also planned to reduce this contention") yet and they're working on it with Huawei. Today they recommend "ordinary 8-GPU machines for prefill, ordinary 64–128-GPU supernodes for decode". There is no 64-128 GPU "Nvidia supernodes". It may be they've solved that part at least internally, and that 100TPS case can improve to ≈8-10K in 128-Chip SuperNodes using DeepInfer with high EP. (I ignore the actual mix for simplicity, directionally it'll be similar). In that case, assuming $10M per NVL72 rack (reasonable current smuggling premium, extrapolated from >$1M per 8-GPU DGX B300 server a few months ago) thus $139K per GPU, and $37K per 950DT (rumored), Blackwells likely remain more economical in the actual V4.1F product regime, but clearly lose in the throughput-maximizing one (eg RL rollouts). The crossover is roughly 6.8K tokens/s/chip. Given the efficiency of Blackwell inference and the competitive advantage of V4.1F being that fast (but still making a ton of dumb turns), it'd make sense for DeepSeek to cover the API with a small Blackwell cluster worth <$1B, ie ≈100 racks (it'll pay for itself in well under a year), and fully pivot to Ascend for post-training inference. They will do training of larger future models on Ascend SuperPoDs/SuperClusters, when available.
