In a rather interesting discovery, a 7nm AMD Radeon Vega GPU with 16GB HBM2 memory has been spotted on a DCU designed by Chinese chipmaker Hygon. Keep in mind that this isn’t the Radeon Instinct, rather an engineering sample of the Kehai light Hygon DCU.
The GPU appears to have four HBM 2 stacks for a total bus width of 4096 bits similar to the Radeon Instinct MI150. It’s unclear whether the GPU is the same as well, but that should be the most likely case. (thermi.com)
The MI150 is based on the Vega 20 GPU with 60CUs and a peak GPU clock of 1725MHz. It’s also based on TSMC’s 7nm FinFET node and offers an impressive 13.3 TFLOPs of FP32 performance. On the memory side, you have a wide, but low clocked (1GHz) 16GB of HBM2 memory, resulting in a total bandwidth of 1024GB/s thanks to the wide 4Gbit bus.
The MI150 is an accelerator like NVIDIA’s Tesla lineup and supports mixed-precision compute with a rated performance of 26.5 TFLOPs (FP16) and 53 TOPs of INT8. It comes with support for PCIe 4.0 and has a total board power of 300W.