The capability of Nvidia’s mystery A800 compute GPU, which is developed for the Chinese market, has been exposed in a relatively brief narrative about overwhelming demand for Nvidia’s high-performance computing hardware in China. According to MyDrivers, the A800 operates at 70% the speed of A100 GPUs while adhering to tight US export requirements that limit the amount of processing power Nvidia can sell.

Nvidia’s A100, now three years old, is a powerhouse: it produces 9.7 FP64/19.5 FP64 Tensor TFLOPS for HPC and up to 624 BF16/FP16 TFLOPS (with sparsity) for AI tasks. Even if the values are reduced by roughly 30%, they are still impressive: 6.8 FP64/13.7 FP64 Tensor TFLOPS and 437 BF16/FP16 (with sparsity).

In terms of compute capabilities, despite ‘castration’ (performance limitations), as MyDrivers puts it, Nvidia’s A800 competes with fully-fledged China-based Biren’s BR104 and BR100 compute GPUs.

Meanwhile, Nvidia’s compute GPUs and CUDA architecture are extensively supported by its customers’ applications, whereas Biren’s CPUs have yet to be embraced. Because of the new regulations, even Biren cannot ship full-fledged computing GPUs to China.

The export limits implemented by the US in October 2021 prohibit the transfer to China of American technologies that enable supercomputers with performance exceeding 100 FP64 PetaFLOPS or 200 FP32 PetaFLOPS in a space of 41,600 cubic feet (1,178 cubic metres) or less. While the export restrictions do not directly limit the performance of each compute GPU sold to a Chinese business, they do limit throughput and scalability.

Following the implementation of the new laws, Nvidia lost the ability to sell its ultra-high-end A100 and H100 compute GPUs to Chinese clients without an export licence, which is difficult to obtain. In order to meet the performance demands of Chinese hyperscalers, the firm produced the A800, a scaled-down version of their A100 GPU. It was unclear how capable this GPU was until now.

As the use of artificial intelligence grows among both consumers and organisations, so does the demand for high-performance technology capable of handling acceptable workloads. Nvidia is one of the primary beneficiaries of the AI megatrend, which is why its GPUs are so popular that even the entry-level A800 is sold out in China.

Biren’s BR100 will be offered in an OAM form-factor and will be capable of using up to 550W of power. The chip supports the company’s unique 8-way BLink technology, which allows up to eight BR100 GPUs to be installed per machine.

The 300W BR104, on the other hand, will come in an FHFL dual-wide PCIe card form-factor and will allow up to 3-way multi-GPU setup. According to EETrend (via VideoCardz), both chips employ a PCIe 5.0 x16 interface with the CXL protocol for accelerators on top.

Biren claims that both of its chips are manufactured utilising TSMC’s 7nm-class fabrication process (without specifying whether N7, N7+, or N7P is used). The bigger BR100 has 77 billion transistors, compared to 54.2 billion in the Nvidia A100, which is also built on one of TSMC’s N7 nodes.

The company also claims that in order to overcome TSMC’s reticle size limitations, it had to use chiplet design and the foundry’s CoWoS 2.5D technology, which is entirely logical given that Nvidia’s A100 was approaching the size of a reticle and the BR100 is expected to be even larger due to its higher transistor count.

Also Read:

Source