On paper, the BR100’s GPU is based on Biren’s own BiLiren architecture, which in turn is manfactured from a 7nm die lithography and 2.5D CoWoS packaging, courtesy of TSMC. Additionally, the GPU also plays house to 77 billion transistors and for memory, it uses 64GB of HBM2e memory, giving it a maximum bandwidth of 2.3Gbps. On top of that, it also supports the latest PCIe Gen5 and CXL interfaces. In regards to the BR100’s performance, Biren says that its GPGPU is capable of delivering as many as 1000 TFLOPs of power in 16-bit floating point operations, and up to 2000 TFLOPs of 8-bit fixed-point compute power. For comparison sake, those speeds puts the Chinese brand’s product right up the back of NVIDIA’s own H100 Hopper GPU, which houses 80 billion transistors but uses a more advanced 4nm process node. Along with its BR100, Biren also announced that it was launching its own software platform for the GPGPU, known as BIRENSUPA. As per the brand’s official statement, its software will support mainstream deep learning frameworks, including Baidu’s own PaddlePaddle. To put it in another way, it is more or less a certainty that the product will be used in the application of artificial intelligence and machine learning. For that matter, Biren says that its “unified” hardware and software architecture is designed for medical imaging, molecular dynamics, and electromagnetic simulations. Alongside the BR100, Biren also launched a smaller GPGPU known as the BR104, which houses half the compute power and memory, is designed for the PCIe form factor, and has TDP of 300W. At the time of its announcement, Biren did not specify any pricing or availability of the BR100 or BR104. (Source: Videocardz, Tweaktown, SCMP)
