Biren Technology has formally introduced its first GPUs designed primarily for artificial intelligence (AI) and high-performance computing (HPC). According to the company, the top-of-the-range BR100 GPU can challenge Nvidia's A100 and even H100 chips in certain workloads, yet its complexity is comparable with that of Nvidia's H100 compute GPU.
Biren's initial family of compute GPUs includes two chips. The BR100 promises up to 256 FP32 TFLOPS or 2 INT8 PetaFLOPS performance, whereas the BR104 is rated for up to 128 FP32 TFLOPS or 1 INT8 PetaFLOPS performance.
|Biren BR104||Biren BR100||Nvidia A100||Nvidia H100|
|Form-Factor||FHFL Card||OAM Module||SXM4||SXM5|
|Transistor Count||?||77 billion||54.2 billion||80 billion|
|FP16 TFLOPS Tensor||?||?||312/624*||1000/2000*|
|BF16 TFLOPS Tensor||?||?||312/624*||1000/2000*|
|INT8 TFLOPS Tensor||?||?||624/1248*||2000/4000*|
* With sparsity
Both chips support the INT8, FP16, BF16, FP32, and TF32+ data formats, so we're not talking about supercomputing formats (e.g., FP64) even though Biren says that its TF32+ format provides higher data precision than traditional TF32. Meanwhile, the BR100 and BR104 offer rather formidable peak performance numbers. In fact, if the company had incorporated GPU-specific functionality (texture units, render back ends, etc.) into its compute GPUs and had designed proper drivers, these chips would have been rather incredible GPUs (at least BR104, which is presumably a single-chip configuration).
In addition to the compute capabilities, Biren's GPUs can also support H.264 video encoding and decoding.
Biren's BR100 will be available in an OAM form-factor and consume up to 550W of power. The chip supports the company's proprietary 8-way BLink technology that allows the installation of up to eight BR100 GPUs per system. In contrast, the 300W BR104 will ship in a FHFL dual-wide PCIe card form-factor and support up to 3-way multi-GPU configuration. Both chips use a PCIe 5.0 x16 interface with the CXL protocol for accelerators on top, reports EETrend (via VideoCardz).
Biren says that both of its chips are made using TSMC's 7nm-class fabrication process (without elaborating whether it uses N7, N7+, or N7P). The larger BR100 packs 77 billion transistors, outweighing the 54.2 billion with the Nvidia A100 that's also made using one of TSMC's N7 nodes. The company also says that to overcome limitations imposed by TSMC's reticle size, it had to use chiplet design and the foundry's CoWoS 2.5D technology, which is completely logical as Nvidia's A100 was approaching the size of a reticle and the BR100 is supposed to be even larger given its higher transistor count.
Given the specs, we can speculate that BR100 basically uses two BR104s, though the developer has not formally confirmed that.
To commercialize its BR100 OAM accelerator, Biren worked with Inspur on an 8-way AI server that will be sampling starting Q4 2022. Baidu and China Mobile will be among the first customers to use Biren's compute GPUs.