While much of the focus this past year has been on China’s pivot towards RISC-V, the reality is that the US has been exerting pressure to limit its use in the Asian country and its companies. Now, China is unleashing its magic with LinxISA, an architecture that directly targets the core of hardware: the ISA, or the set of instructions that dictates how software communicates with a given chip. Could this be the ultimate solution to bypass US restrictions?
Presented by Huawei from China, LinxISA is an open-source architecture published on GitHub with a highly ambitious goal: to unify CPUs, GPUs, and Artificial Intelligence accelerators under a single technical foundation.
LinxISA: Huawei Enters the Race with an ISA Aiming to Unify AI Chips in One Stroke
The truly interesting aspect of LinxISA lies not just in it being another instruction architecture, but in its unique approach. Huawei defines LinxISA in its repository as a block-structured instruction architecture oriented towards high-performance computing, adopting a philosophy distinct from that of a traditional scalar ISA.
LinxISA’s design revolves around these instruction blocks, which are executed as atomic units. Each block has explicit boundaries, control flow integrity rules, and a state model divided into two layers: global state and local state within the block itself. This structure allows for better execution organization and paves the way for models better suited for parallel workloads.
This is where the most striking aspect of the Chinese company’s development comes into play. LinxISA doesn’t seem to be designed solely for classic CPUs; the project includes vector and tile extensions, geared towards SIMD-type parallelism and workloads closer to GPUs, matrix accelerators, and Artificial Intelligence. In simple terms, the idea is to reduce the gap between general computation, graphics, and specialized accelerators. In essence, one ISA for all, and all designed to work with that ISA.
Creating an ISA for Everything and Everyone
According to information emerging from China, LinxISA is envisioned as potentially one of the first recent open-source ISAs with a genuine ambition to unify CPUs, GPUs, and AI accelerators from the ground up in its architectural design. This isn’t just about adding new instructions; it’s about enabling different types of processors to share a common execution logic.
The open-source component is also crucial. The LinxISA/linx-isa repository is publicly available on GitHub under the BSD 3-Clause license, allowing anyone to download it and begin designing their own hardware. Furthermore, it goes beyond a simple specification: the project includes components such as LLVM, QEMU, the Linux kernel, RTL LinxCore, glibc, musl, validation tools, documentation, workloads, and an ISA specification marked as v0.56. It’s akin to the “CUDA of ISAs,” where Huawei aims to encourage a broad range of manufacturers to design with LinxISA, rather than opting for RISC-V.
Indeed, Huawei reportedly filed trademark applications related to LinxISA in 2019 and 2022. While the repository is now available, the code is published, and the architecture is documented, there hasn’t been a direct, official announcement from Huawei publicly confirming all project details in the reviewed information. This seems to be their typical approach: launch it and let everyone pay attention.
Consequently, this represents a direct challenge to RISC-V and its consortium. With LinxISA, the US may no longer need to exert as much effort to control RISC-V usage in China, as China now has its own alternative. The question remains: will US companies eventually adopt a Chinese ISA? If so, it’s highly probable that products created with it, regardless of the hardware, might face import restrictions into the US. Only time will tell.
