DeepSeek's open-source kernels reveal Huawei Ascend 950's 32-core architecture and frequency gains
teortaxesTex · x · 2026-10-03
- DeepSeek's open-sourced DeepGEMM, FlashMLA, TileKernels and DeepEP let analysts reverse-engineer Huawei's Ascend 950 beyond spec sheets.
- Each AI Core packs 1 Cube Core + 2 Vector Cores: 32 AI Cores, 32 Cube Cores and 64 Vector Cores total. Cube handles dense matmul; Vector handles dequantization, Softmax and more.
- skip-scale optimization cuts FixPipe transfers, vector work and memory movement; DeepSeek notes lower compute/data-movement power lets the chip clock higher, improving performance.
- The analysis traces the full path from matrix compute and memory movement to multi-NPU communication, showing open-source kernels as a new lens on China's AI silicon.
Related event: DeepSeek Open-Source Kernels Reveal Custom Chip and Ascend 950 Details(2 posts)→
More from Infra
- Dev submits PR adding periodic checkpoints to GKE sandboxes so agents don't lose progress — yoimnotkesku · 2026-11-19
- AWS Lambda Adds GPU Support via Managed Instances for Inference and Media Workloads — DavidWells · 2026-11-19
- Nvidia Allegedly Made GPUs Artificially Scarce to Keep Prices Sky-High — GaryMarcus · 2026-11-19
- Leak Hints at Classified 1M-GPU Training Cluster in China That Skips HBM — teortaxesTex · 2026-11-19
- Tencent leaned on Japanese neocloud for AI compute; DeepSeek, Zhipu best positioned — teortaxesTex · 2026-11-19
- Open-source prefill engine for GLM-5.3-Flash beats vLLM by 3.2x on 4x B200 — CShorten30 · 2026-11-19