Dongfang Suanxin with 14nm chip without HBM memory

Shanghai-based Chinese startup Dongfang Suanxin has officially broken the silence and unveiled its advanced AI processor, the DF1000. The chip is a direct response to tough US export restrictions that deny China access to cutting-edge lithography machines and high-speed HBM (High Bandwidth Memory).

The DF1000 is built on a mature and proven 14nm process. The company claims that its new silicon can deliver up to 520 teraflops (TFLOPS) of computational power at BF16 precision, the standard for learning large language models. That's about 52.6 % of performance on Nvidia's flagship H100 and H200, but the key advantage lies elsewhere.

Using software-defined computing and 3D near-memory computing, the new chip achieves a memory bandwidth of 6.4 TB/s. This is almost twice as much as Nvidia's H100 and about 33 % more than the H200. Using a special vertical interconnect, classic DRAM memory is stacked directly above the processor circuit, drastically shortening the data path and eliminating the need for prohibitive HBM memory.

While these specs are impressive on paper, the reality is a bit different. The real challenge for Dongfang Suanxin will not just be the silicon itself, but the maturity of their software, driver stability, and the efficiency of connecting multiple cards in data centers.

The company is already planning successors to the DF1000 AI processor. The DF2000 model is expected to be launched at the end of 2026, which is expected to double the performance and directly compete with the H200 card, and the DF3000 model is expected to be launched at the end of 2027. The first shipments of the basic DF1000 model, which the manufacturer claims is ready for mass production and relies entirely on the domestic Chinese supply chain, are expected to be delivered to customers before the end of this year.