Russia is preparing a new line of specialized artificial intelligence chips that are intended to be an alternative to Nvidia accelerators. Baikal Electronics plans to bring these solutions to the market by 2029–2030. The main difference of the project is the declared compatibility with the Nvidia CUDA ecosystem, that is, with the software environment for which many AI solutions have already been written. This is reported by CNews.
The new line is expected to include two main chips – Baikal-AI-E1000 and Baikal-AI-D1000.
Baikal-AI-E1000 is aimed at the industrial segment. It is planned to be used in automation, transport, and monitoring systems. The chip's power consumption should not exceed 30 W, and the frequency is declared up to 2 GHz. In terms of positioning, this solution is close to Nvidia Jetson.
The second chip – Baikal-AI-D1000 – should become a full-fledged server accelerator for data centers and artificial intelligence tasks. It is declared to have a performance of 1000 TFLOPS in FP8 format and 500 TFLOPS in FP16 format. The memory capacity can be up to 64 GB. The expected price of Baikal-AI-D1000 is about 10,000 dollars.
The main goal of the project is to create a sovereign AI infrastructure and reduce dependence on foreign suppliers, primarily Nvidia, which currently dominates the accelerator market.
CUDA support should remove one of the main barriers for developers: if the new chip can work with the already accumulated software stack, the transition to a Russian solution may become easier.
