Ep. 017 - DeepSeek V4 and Huawei Ascend NPU Performance (InferenceX) | Kimbo Chen, Cam Quilici, Bryan Shan, Jordan Nanos
SemiAnalysis WeeklyShow Notes
DeepSeek V4 claims a 100x KVcache reduction versus a standard MoE model, hitting 1M context length through compressed sparse attention and heavily compressed attention. Kimbo (@Kimbochen), Cam Quilici (@noslawextratost), Bryan Shan join Jordan Nanos (@JordanNanos) to break down what changed from V3, why the new MHC dimension tripped up NVIDIA on day zero, and how Mega MoE fuses compute and communication into a single kernel. The vLLM versus SGLang NDA access gap and the Huawei day zero optimization guide circulating on Twitter.The crew walks through the InferenceX article on going from day zero to day 43 support and what that grind actually looks like across different hardware. Subscribe for weekly semiconductor and AI infrastructure analysis from the SemiAnalysis team.
Referenced:DeepSeekV4 1.6T Day 0 to Day 43 Performance Over Time - GB300 NVL72, Huawei, MI355X, B200: https://newsletter.semianalysis.com/p/deepseekv4-16t-day-0-to-day-43-performanceChapters:
(00:00) DeepSeek V4 vs V3 changes(01:00) Sparse attention and KV cache reduction(03:04) Day zero runtime support challenges(05:34) What Mega MoE actually is(08:38) Downsides of fusing kernels(10:25) MegaKernel benchmark claims(12:59) AMD FP4 optimization gains(15:14) Compounding step by step improvements(17:59) vLLM versus SGLang competition(19:34) Open source vs vendor libraries
More from SemiAnalysis Weekly
Brad Gerstner put our 43 gigawatt forecast for 2027 on All-In, then again in his BG2 sit-down with Jensen, and called it too aggressive. The consensus counter sits near 25 GW, built on moratorium headlines and grid connection fears. Jordan Nanos (@JordanNanos) sits down with Jeremie Ontiveros (@JeremieEO) and Reyk Knuhtsen (@robotknower), who have been tracking the same capacity curve since March 2024, and it has barely moved across every model revision. So they aren't cutting it. They're raising it.
GLM 5.3 only attends to the tokens that matter. It still has to keep every one of them.
H100 rental prices went up, on a GPU that should be sliding down its depreciation curve. That curve is what lenders underwrite. Daniel Nishball (@dnishball) kicks off our new Compute Capital Markets group by following the money through Nvidia's $497B backstops, ClusterMAX as a credit signal, and whether each token is profitable. Jordan Nanos (@JordanNanos) brings the rental data. AI debt is about to pass auto loans.
Pay Anthropic $200 a month and you can pull $12,000 of Opus 5.5 tokens at API pricing. Max Kan (@maxkan) and Andrew Megalaa bought every plan, OpenAI, Meta, xAI, MiniMax, Moonshot, Z.ai, Cursor, and measured every token type to find the ceiling. The catch isn't subsidy. Anthropic still books 50 to 60% gross margins on these plans, which means the API itself is the price gouge.
Jordan (@JordanNanos) and Abhilash (@AbhilashJain17) discuss the clients SemiAnalysis consulting serves, custom projects like the inference simulator, how client work turns into research products, neocloud due diligence, and why SLAs decide who gets funded. They also discuss Feynman specs, de-specced memory, and a 20-trillion-parameter OpenAI model.