Ascend 950DT hits 1/5-1/7 of GB200 decode throughput on DeepSeek-V4.1, but EP optimization may close the gap

teortaxesTex · x · 2026-10-21

A detailed analysis compares Huawei Ascend 950DT vs NVIDIA GB200 running DeepSeek-V4.1-Flash decode:

• Per the Sep 30 CANN benchmark, 950DT delivers 5K tokens/s/chip at 100 tps/user with 131K context, vs 25.5K on GB200 with DeepInfer (EP64-128/DP64-128) — a 5x gap; at 230 tps it's 20K vs 2.7K, or 7.4x.
• Key caveat: DeepSeek's DeepEP-Ascend docs (3 weeks ago) state EP64-128 is not yet optimized — combine has local reduction overhead and HBM contention with URMA — with Huawei firmware fixes planned.
• DeepSeek now recommends ordinary 8-GPU machines for prefill and 64–128-GPU supernodes for decode, a topology with no NVIDIA equivalent; with high-EP DeepInfer on 128-chip supernodes, the 100-TPS point could rise to 8-10K.
• Economics: at $139K/Blackwell vs $37K/950DT, GB200 needs 3.76x per-chip throughput to break even on acquisition cost. Today GB200 has 33% more throughput per dollar; an optimized 950DT at 8K-10K would flip that to 18-47% in Ascend's favor (excluding networking, cooling, power, etc.).

Takeaway: the headline 5-7x gap is partly a software-stack artifact and may shrink or reverse on per-dollar economics.

Original post →

More from Infra

• 275 messages in 20 days: oss-sec list floods with Apache CVEs — DanielLockyer · 2026-10-21
• teortaxesTex challenges robot-token thesis: supply chain can't scale by 2027 — teortaxesTex · 2026-10-21
• 2.5T USD ARR needed by 2027 to justify 50GW of AI power, even with 75% price crash — bookwormengr · 2026-10-21
• Nokia Bell Labs laser linewidth improved 50–60 dB, chart made interactive with Claude — jwt0625 · 2026-10-21
• Do 50GW of AI compute need $10T ARR? A DeepSeek-pricing back-of-envelope debate — teortaxesTex · 2026-10-21
• One of the best things about the AI era: infra work is interesting again — sloppenheimer · 2026-10-21

Infra channel →