Baikal Electronics demonstrated a working prototype of its own GPGPU core, implemented on an FPGA. The company showcased its operation in computer vision tasks, including real-time human recognition. The development is intended to become the basis of the future Baikal-AI-E1 processor, which the manufacturer positions as a functional analogue of Nvidia Jetson Orin.
GPGPU is a computational architecture focused on parallel processing of large volumes of data. Such solutions are used not only in graphics but also in computer vision, machine learning, and other computations where many operations need to be processed simultaneously.
The Baikal-AI prototype was implemented on an FPGA - a field-programmable gate array. This approach allows for refining the architecture of the future processor before the release of its own silicon chip. The presented version uses its own GPGPU block, a second-level cache, a control processor based on the RISC-V architecture, and a software stack.
The company also claims compatibility of the developed GPGPU with CUDA. Existing computer vision models are already running on the prototype, and the next step will be to transfer the development from FPGA to its own silicon chip.
Baikal-AI-E1 is expected to be the first processor based on this architecture. Baikal Electronics calls it a functional analogue of Nvidia Jetson Orin, designed for local data processing. This approach allows computations to be performed directly on the device, without the mandatory transfer of data to a remote server.
Among the possible application areas, the company names computer vision systems, unmanned aerial vehicles, robotics, industrial quality control, and smart cameras.
However, the version shown now is a GPGPU prototype on an FPGA, not a ready-made serial AI processor. The real characteristics of Baikal-AI-E1 in terms of performance, power consumption, and die area can be evaluated after the creation of the silicon version.