Chinese GPU maker: Biren BR100 has 77 billion transistors, 64GB HBM2e
China's new Biren BR100 GPU: 77 billion transistors, 7nm process node, 64GB HBM2e memory @ 2.3TB/sec memory bandwidth, PCIe 5.0.
Chinese company Biren Technology has just unveiled its new GPGPU architecture for next-gen computing at its recent Biren Explore Summit 2022 event.
Biren's new BR100 GPU is built on the 7nm process node at TSMC using their TSMC 2.5D CoWoS packaging technology, featuring 77 billion transistors, 300MB of on-chip cache, and 64GB of super-fast HBM2e memory that pumps up to 2.3TB/sec of memory bandwidth. We have up to 256 TFLOPs of FP32 compute performance, with the GPU supporting both PCIe 5.0 and CXL interfaces.
The new BR100 GPGPU is based on Biren's own BiLiren architecture, but with the BR100 the company is also introducing the whole family of OAM-compatible cards, as well as the PCIe-based BR104 GPU. Biren is also unveiling their own software platform alongside the BR100 GPGPU launch, which is called BIRENSUPA, with support for mainstream deep learning frameworks like Baidu's PaddlePaddle.
- Read more: Chinese GPU makers should have 7nm, 5nm GPUs on shelves in 2022
- Read more: Moore Threads MTT S60, MTS2000 GPU: first China domestic GPU arch
- Read more: Jing Jiawei JM9: Chinese GPU maker keeps up with GeForce GTX 1080
- Read more: Innosilicon Type-B graphics card: Chinese GPU has 32GB GDDR6X memory
If we compare the Biren BR100 GPGPU up against the NVIDIA H100 "Hopper" GPU, Biren isn't too far behind: 77 billion transistors on the BR100 versus 80 billion transistors on the H100. NVIDIA is building its Hopper H100 GPUs on TSMC's fresh new 4nm process node, too.
Biren also introduced the new BR104 GPU, which is half as powerful as the BR100. The big difference here is that the BR104 is a monolithic GPU design, compared to the BR100 using a chiplet GPU design... as well as the Biren BR104 GPU only arrives in the regular PCIe form factor, with a 300W TDP.
Biren says that its new BR100 GPGPU with 77 billion transistors is capable of mimicking the human brain nerve cells, while the BR100 will be used for DNN and AI tasks... which will slowly reduce China's reliance on NVIDIA's GPU for AI-powered tasks.
We don't know pricing just yet, but that is going to be very interesting to find out in the coming weeks and months.