BIREN
Chinese company Biren Technology introduced the BR100 GPU, which offers performance of up to 1024 Tflops in 16-bit floating point operations and up to 2048 Tflops in 8-bit integer operations. The solution is intended for use as part of specialized high-performance computing (HPC) systems.
The BR100 GPU is the company's flagship solution, using TSMC's 7nm manufacturing process and 2.5D CoWoS packaging technology. The chip uses the BiLiren architecture and has 77 billion transistors on board.
Solutions based on it can offer up to 64GB of high-speed HBM2e memory with a maximum transfer rate of 2.3TB/s, plus 300MB of dedicated cache. For GPGPU, support for PCIe 5.0 and CXL interfaces is declared.
The performance of the BR100 is superior to that of the NVIDIA Ampere A100, but about 2.0-2.5 times slower than solutions based on the NVIDIA Hopper architecture. Maximum superiority over NVIDIA Ampere is declared when working with 32-bit floating point numbers - in this case, BR100 can deliver 256 Tflops versus 19.5 Tflops performance for NVIDIA solution.
In addition to the BR100, the Chinese company also introduced the BR104 graphics chip. It offers half the specs and about half the performance of the BR100, and can be used in PCIe expansion cards with power consumption up to 300W.
Biren Technology has also introduced the BIRENSUPA software platform for its products. It supports several popular machine learning frameworks such as Baidu's PaddlePaddle.
mundophone
No comments:
Post a Comment