Suanmiao (算苗科技) announces tape-out of its A4E 3D TokenPU chip, a dedicated inference processor for l...
The AMW Read
Novelty 2: Suanmiao is a new entrant in the inference silicon space but follows known 3D-stacking pattern; significance 2: updates the inference-chip competitive landscape within China's sovereign supply chain push, with cross-substrate compute and geopolitical implications.
Suanmiao (算苗科技) announces tape-out of its A4E 3D TokenPU chip, a dedicated inference processor for large models built on domestic supply chain. The chip uses 8-layer vertically stacked memory dies atop compute logic, connected through TSV and bump technology, claiming 16 TB/s memory bandwidth. Suanmiao employs a Tile-Native architecture that treats data tiles as the fundamental unit for movement, storage, and computation, aiming to reduce the energy and cost overhead of data movement that the company says accounts for 80% of power and 70% of cost in typical inference workloads. The chip is built on RISC-V cores with self-developed IP and software stack compatible with LLVM and Triton. Suanmiao states it has worked with leading Chinese foundation model labs for nearly a year, co-designing the architecture against real inference requirements.
Why it matters: Suanmiao's tape-out exemplifies the sovereign AI silicon push within China, leveraging 3D hybrid stacking to overcome memory bandwidth constraints without relying on advanced process nodes. This fits the pattern of inference-optimized architectures emerging as the market pivots from training to inference—Deloitte projects over 80% of future AI compute will be inference-side. Suanmiao is not a general-purpose GPU competitor; it is a purpose-built inference accelerator that targets the same market opening that Groq, Cerebras, and others address outside China. The company's ability to combine domestic front-end design with domestic packaging and manufacturing partners signals maturation of China's alternative chip ecosystem, though volume production and real-world LLM throughput data remain unverified. The involvement of state-backed investors (China Development Bank, Beijing Shunxi) alongside market VCs (Source Code, Lenovo Capital) places this squarely in China's sovereign AI infrastructure buildout.
Expert take: Suanmiao's core team includes engineers with prior experience in high-throughput 3D-stacked chips, making this more credible than a pure paper architecture. The Tile-Native approach mirrors some ideas in Cramer's sparse-attention and memory-centric design patterns, though Suanmiao is not a known player in Western benchmarks or software ecosystems. The strategic bet is that Chinese hyperscalers will prioritize domestic inference silicon as US export controls tighten, creating a captive market large enough to support a dedicated inference chip vendor. The open question is whether Suanmiao can deliver competitive tokens-per-second and latency on models like Qwen, DeepSeek, or Ernie without Nvidia's CUDA ecosystem—its LLVM/Triton compatibility is a necessary but insufficient condition for developer adoption.