The Symbiosis of Huawei and DeepSeek: Ascend 950PR Real-World Test | Huawei Ascend | DeepSeek | A...

Original Media · 17m · 中文 · YouTube · Best Partners TV

TLDR

This episode analyzes the symbiotic relationship between Huawei and DeepSeek, where Huawei provides hardware (Ascend 950 series) and software stack (CANN), while DeepSeek contributes model workloads and kernel engineering. The collaboration is evidenced by DeepSeek's Sept 30 open-source release of six repositories with Ascend support. The discussion covers hardware microbenchmarks revealing chip architecture, performance comparisons showing Ascend 950PR nearing theoretical limits but lagging absolute performance, software innovations like TileLang and DeepEP-Ascend enabling efficient kernel compilation and communication-computation overlap, and ecosystem gaps limiting broader adoption beyond Huawei-DeepSeek supernodes.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

Ascend 950PR's native torch.matmul (aclnnMatmul) reaches 99.7% theoretical peak without DeepSeek's custom kernels, showing Huawei's own engineering maturity.

•

The chip contains 5-7 AICPUs (likely Kirin 9000S middle cores) that handle KV metadata and scheduling, not just device control.

•

DeepSeek's FlashMLA KV cache layout had to be redesigned for Ascend: co-locating quantized values and scaling factors to reduce memory transactions.

•

SemiAnalysis calls this "Jensen Huang's nightmare becoming reality" — the ecosystem gap is closing faster than expected due to co-design.

Key takeaways

•

Huawei and DeepSeek have a mutual dependency: Huawei lacks a frontier lab to validate chips; DeepSeek lacks compute access.

•

Ascend 950PR uses extensive binning (multiple SKUs) due to die defects and HBM stacking yield issues, unlike NVIDIA's compute-only binning.

•

Microbenchmarks (pointer chasing, memory bandwidth) reverse-engineer chip topology: 2 dies, 6 subsystems (0,2,4 / 1,3,5), each with 1 Cube + 2 Vector cores, plus 5-7 AICPUs per chip.

•

Ascend 950PR achieves 99.7% of theoretical GEMM peak (vs DeepGEMM's 99%+), but absolute performance sits between A100 and MI300X; memory bandwidth utilization exceeds H100 for small messages.

•

TileLang's Ascend backend is merged upstream, providing native compilation/scheduling/runtime, reducing kernel development barrier; AutoSchedule cuts small-operator latency up to 40%.

•

DeepEP-Ascend ports V2.5 to URMA/HCCL; communication-computation overlap is constrained by shared Vector units and memory bandwidth, requiring careful resource balancing (as seen in Kimi K3 recipe).

•

DeepSeek-V4.1-Flash uses speculative decoding (DSpark) and Engram (async token ID loading), with profiling showing 88.8% stream utilization but only 72.6% device time due to overlap.

•

Ecosystem gaps remain: x86 host support is broken (hardcoded ARM headers, no CI), optimizations target supernodes, making deployment outside Huawei-DeepSeek config difficult.

Notable quotes

“单看这个动作 像是一次普通的开源提交 但是放到中美算力的背景里 它更像一个信号”

▸ Notable quotes · 0:19

“华为出硬件和原生软件栈 DeepSeek 出工作负载 还有让这些硬件真正服务自己模型的那部分工程”

▸ Notable quotes · 0:05

“硬件抬高了上限 但软件才能让你摸到那个上限”

▸ Notable quotes · 6:53

“所谓共生,重点不在算子本身 而在于让应用的其余部分、以及支撑这些部分的工作 变得更高效、更容易复用”

▸ Notable quotes · 7:28

“黄仁勋的噩梦 正开始成为现实”

▸ Notable quotes · 16:55

“如果这套协同只在 DeepSeek 和华为之间跑得通 那它到底是国产算力生态的地基 还是一条定制化的专线呢?”

▸ Notable quotes · 16:28

Chapters

•

0:00 — Introduction: The Symbiosis Signal: DeepSeek's Sept 30 release of six repos with Ascend support signals a deepening Huawei-DeepSeek partnership, framed by SemiAnalysis as a mutual dependency: Huawei needs a frontier workload to validate chips, DeepSeek needs compute.

•

1:36 — Hardware Deep Dive: Ascend 950PR Architecture & Binning: Details on 950PR variants due to die defects and HBM yield; microbenchmarks (pointer chasing) reveal 2 dies, 6 subsystems (1 Cube + 2 Vector each), and 5-7 AICPUs (Kirin-derived) per chip handling KV metadata and scheduling.

•

4:17 — Performance Analysis: GEMM, Memory Bandwidth & TCO: Native aclnnMatmul hits 99.7% theoretical peak; absolute performance between A100 and MI300X. Memory bandwidth utilization exceeds H100 for small messages. Comparison complicated by NVIDIA power limits; TCO advantage for China due to cheaper electricity.

•

7:02 — Software Co-Design: TileLang, DeepEP-Ascend & Communication Overlap: TileLang Ascend backend merged upstream with AutoSchedule (up to 40% latency reduction). DeepEP-Ascend uses URMA/HCCL; overlap limited by shared Vector units and bandwidth, requiring resource balancing (Kimi K3 limits Vector usage).

•

12:07 — Model-Level Innovations: Speculative Decoding & Engram: DeepSeek-V4.1-Flash uses DSpark speculative decoding and Engram (async token ID loading). Profiling shows 88.8% stream utilization, 72.6% device time due to overlap; attention rewritten for Ascend's L0C accumulator constraints.

•

15:05 — Ecosystem Gaps & Deployment Reality: x86 host support broken (ARM hardcoded headers, no CI); optimizations target supernodes only. Engineering debt surfaces outside Huawei-DeepSeek config, limiting broader adoption.

•

16:55 — Conclusion: Foundation or Custom Line?: Ecosystem gap closing ("Jensen's nightmare"), but key question remains: will this co-design become a general foundation for Chinese AI industry, or stay a custom Huawei-DeepSeek line? Answer depends on community adoption and upstream contributions.

Speakers

•

大飞 (Da Fei) — Host/Analyst

⋮