Skip to main content

算苗科技（北京）有限公司

Category: AI Chips / Semiconductors

Chinese AI chip startup pioneering 3D TokenPU architecture for large-model inference, using domestic supply chain and 3D hybrid stacking technology. 算苗科技（北京）有限公司 was founded in 2022. The company is led by 汪福全 (Wang Fuquan). Based in Beijing, China. Team size: 101-500. Total funding raised: ¥1.0B (~$137M). Latest round: Pre-Series A. Key investors include 源码资本 (Source Code Capital), 石溪资本 (Stony Creek Capital), 联想创投 (Lenovo Capital), 襄禾资本 (Xianghe Capital), 国开金融 (CDB Capital), 北京顺禧 (Beijing Shunxi Fund).

Founded
2022
Headquarters
Beijing, China
Team size
101-500
Total funding
¥1.0B (~$137M)

Value proposition

Proprietary 3D TokenPU architecture that vertically stacks memory on compute logic, delivering 16 TB/s+ memory bandwidth using mature domestic process nodes (12nm), achieving 1.26-2.19x the inference throughput of NVIDIA H200 on Llama/Mixtral models at significantly lower TCO.

Products and solutions

3D TokenPU A4E (first-gen inference accelerator, taped out June 2026, 16 TB/s memory bandwidth, 8-layer DRAM stacked on compute logic via TSV/bump), A4S (second-gen, definition complete, expected tape-out Feb 2027), Tile-Native software-hardware co-design ecosystem with LLVM/Triton compiler toolchain support

Unique value

World's first 3D TokenPU purpose-built for LLM inference; breaks the memory wall by stacking 8 DRAM layers directly on compute logic via TSV; achieves superior inference performance using mature domestic 12nm process vs NVIDIA's 4nm; fully domestic supply chain (RISC-V based, Chinese foundry, domestic DRAM partners).

Target customer

Major LLM companies (top-tier model providers), cloud service providers, AI inference infrastructure operators

Industries served

AI Cloud Computing, Large Language Model Inference, Data Center AI Acceleration

Technology advantage

Proven 3D hybrid bonding mass production capability (>10,000 wafers); fully domestic supply chain (RISC-V architecture, CXMT/SMIC/GigaDevice partners); Tile-Native software-hardware co-design; team's prior crypto chip JASMINER X4 achieved 20x power efficiency vs NVIDIA 7nm GPUs; 12B+ RMB cumulative 3D chip revenue from prior venture.

How they differentiate

Unlike GPU-centric competitors (NVIDIA, Biren, Moore Threads) that pursue general-purpose compute, Sunmmio's 3D TokenPU is a dedicated ASIC for LLM inference. The 3D hybrid stacking approach eliminates the memory bandwidth bottleneck by shortening data travel distance from millimeters to micrometers. The team is the world's first to achieve commercial-scale 3D stacked chip mass production with >10,000 wafers shipped, using fully domestic supply chain (CXMT DRAM, SMIC foundry, GigaDevice).

Main competitors

NVIDIA (H200/B200 for inference), 壁仞科技 (Biren Technology), 寒武纪 (Cambricon Technologies)

Key partnerships

长鑫存储 (CXMT - 3D DRAM partner), 中芯国际 (SMIC - foundry partner), 兆易创新 (GigaDevice - 3D custom DRAM partner), Domestic RISC-V ecosystem partners

Notable customers

Major Chinese LLM companies (names not publicly disclosed, described as "top-tier model providers" with ~1 year of deep co-development)

Major milestones

2022-11: Company founded, 2025-12: First funding round (Angel/Pre-A) led by Source Code Capital and Stony Creek Capital, 2026-02: Pre-A1 round led by Xianghe Capital with CDB Capital and Beijing Shunxi, 2026-06-15: First-gen 3D TokenPU A4E successfully taped out using fully domestic supply chain, 2027-02 (target): Second-gen A4S tape-out

Growth metrics

~190 R&D staff (total ~230+ employees); 80%+ R&D staff; cumulative 3D chip revenue from predecessor company >¥12B RMB; first-gen chip A4E taped out June 2026

Market positioning

Pioneer in 3D-stacked ASIC for LLM inference in China; positioned as a domestic alternative to NVIDIA GPUs for inference workloads; backed by both market-driven VCs and national strategic capital; targets the rapidly growing inference market (projected 80%+ of future AI compute demand).

Geographic focus

China (primary), with stated ambition for global markets

Patents and IP

Multiple patent applications including DRAM access methods, crossbar switch/SoC access control methods, and 3D chip architecture innovations (verified via Tianyancha).

About 汪福全 (Wang Fuquan)

Ex-Chinese Academy of Sciences (Institute of Acoustics PhD, Institute of Computing Technology postdoc under Loongson chief scientist Hu Weiwu); Founder of Zhongke Shenglong (JASMINER), a crypto compute chip company that achieved 12B RMB in 3D chip revenue; 15+ years in semiconductor and high-throughput computing.

Official website: