Silicon Vanguard: Ranking China's Domestic Chip Leaders (www.machineyearning.io)

Edgerunner Ventures published "Silicon Vanguard," a 13,000-word deep-dive and v0.1 dataset that ranks 16 Chinese fabless accelerator (GPU/ASIC) designers and lays out the hardware powering China’s push for a sovereign AI stack. The report compiles primary-source specs from Chinese filings and archives, groups firms into tiers, and highlights rapid progress not just in raw peak performance but—critically—in energy-compute efficiency, sparse computing, and near-memory/analog approaches. It also stresses ecosystem moves that blunt export controls: Huawei HiSilicon plans to open-source a Unified Cache Manager (UCM) that shards model memory across HBM, DRAM, and SSDs (claimed latency reductions up to 90% and throughput improvements up to 22x), and industry-wide adoption of FP8 (UE8M0) and third‑party baselines (DeepSeek, CAICT) to certify “good enough” inference for large models. Technically notable claims include Moffett AI’s first‑gen chips outperforming NVIDIA H100 in MLPerf Inference (>1.6x throughput at ~3x lower energy, ~5x tokens/joule), growing FP8 support in new SKUs, and improved CUDA-compatibility via transcompilers (Cambricon’s Qimeng‑XPiler reports ~95% translation accuracy with <5 hours debugging). The analysis argues energy-compute optimization—not absolute peak FLOPS—will determine deployment scale and sovereignty, and warns that containment via export controls is likely to accelerate native alternatives. For AI/ML practitioners and policymakers, the takeaway is clear: China’s hardware and software stack is maturing fast enough to be operationally competitive, forcing a strategic shift from restriction to platform-level competition.