Process constraints drive architectural innovation
Traditional chip scaling faces growing hurdles for domestic players amid restricted access to advanced nodes, while AI compute demand soars.
Multiple top firms are now pushing 3D integration and advanced packaging as systemic architectural solutions.
DF1000 debuts, others follow suit
On July 13, Dongfang Suanxin launched its first large‑compute chip DF1000, built on 14nm and delivering 520 TFLOPS peak performance.
Key features: software‑defined silicon for dynamic resource allocation, and 3D‑stacked near‑memory computing with 6.4 TB/s bandwidth, alleviating the memory wall.
The company also claims a preliminary fully domestic supply chain and is fostering a 3D chip industrial cluster.
Other recent moves: Sugon built China's first domestic 100,000‑card AI cluster (July 10); Huawei released "Tao's Law V2" (July 3), proposing architecture to bridge process gaps; Suanmiao's 3D TokenPU A4E taped out in mid‑June. The flurry confirms this path is now a shared industry strategy.
Competition shifts to system‑level capabilities
Analysts agree: competition is moving from node technology to packaging integration and full‑stack efficiency. Guosen Securities cites "chip + HBM + interconnect + server + compiler + inference engine + model adaptation" as the new metric.
Dongwu Securities notes Tao's Law V2 is in engineering verification, with value expected to spread to packaging equipment and high‑speed interfaces.
In capital markets, the sector was weak last week, but short‑sellers added to 15 concept stocks. Unigroup Guoxin saw net buying over RMB 500M; TCL Technology, Muxi, and Kunlun Tech each over RMB 100M.
Muxi's fully domestic GPUs have shipped over 55,000 units by end‑2025; Jingjia Micro's products are deployed in government, telecom, and power.
Near‑term, users won't directly see this chip, but if the tech matures in 2‑3 years, it could reshape cloud compute supply. Expanded domestic AI autonomy may lower service costs and improve stability—but hinges on successful engineering and commercialization, still to be proven.