A Chinese Startup Goes Beyond Nanometers, New Weapons to Chase NVIDIA in AI
The Chinese startup Dongfang Suanxin, also known as Shanghai Oriental Computing Technology, aims to compete in the artificial intelligence chip sector by pursuing an alternative approach to the traditional race towards increasingly advanced manufacturing processes.
On one hand, like many other Chinese companies, it is unable to access the most advanced manufacturing processes from TSMC and other foundries; therefore, this new ambitious entity must make virtue out of necessity, focusing on new architectures and software optimizations.
Dongfang Suanxin's project revolves around two main elements: the so-called software-defined computing and a 3D-stacked near-memory architecture. The goal is to circumvent limitations by intervening in how data and computational resources are managed, according to an article from the South China Morning Post.
The concept of software-defined computing entails dynamic management of the internal resources of the chip, allowing real-time modifications to data flow and organization of computing units based on workload. The aim is to better adapt hardware to various applications, from AI model training operations to inference.
The second technology concerns three-dimensional memory integration. In a near-memory architecture, memory layers are physically positioned closer to computation components, reducing the distances traveled by data compared to traditional configurations. This approach can help reduce latencies and energy consumption while improving available bandwidth.
During the presentation event, Dongfang Suanxin showcased its first AI accelerator, dubbed DF1000. The chip is manufactured using a 14-nanometer process and, according to specifications provided by the company, achieves a computing capacity of 520 teraflops in BF16 format, a numerical representation widely used in training AI models. The processor is also expected to provide memory bandwidth of 6.4 TB/s and chip-to-chip communication speed of 900 GB/s. The company claims that the DF1000 is already ready for mass production, with the first shipments expected by the end of the year.
Alongside the launch of the new accelerator, Dongfang Suanxin outlined an ambitious roadmap. The successor DF2000, anticipated for the fourth quarter of 2026, is expected to double the capabilities of the current model with a stated goal of surpassing the performance of NVIDIA H200. Meanwhile, the DF3000, designed for further performance improvement, is expected in 2027 to compete with the NVIDIA B300 generation, also known as Blackwell Ultra.
Surrounding the DF1000, the company is building a complete ecosystem that includes the general-purpose accelerator module Dianfeng, the TY64 supernode distributed computing system, and the integrated appliance QY100. To complete the platform, CAAP, an open software stack designed to support major AI development frameworks and enable custom programming of operators, supernodes, and computing clusters, has also been introduced.
Dongfang Suanxin's strategy reflects that of the Chinese semiconductor industry, which is seeking alternative solutions to mere reduction of nanometers. Huawei, with its recently updated Tau Scaling Law, is also looking at different parameters to increase the power of its chips.
Founder Wei Shaojun of Dongfang Suanxin, who is also the vice president of the China Semiconductor Industry Association and a professor at the School of Integrated Circuits at Tsinghua University, emphasized the need for China to develop an independent supply chain while also acknowledging some limitations of this approach.
The three-dimensional stacking technology of chips can indeed pose problems in yield productivity, while the limited availability of advanced manufacturing processes remains one of the main obstacles to reaching performance levels comparable to those of world leaders.