A new Chinese GPU by a firm called Biren Technology will reportedly take on Nvidia's next-gen 5nm chips. The GPU isn't gaming-focused and is made for HPC work. The expected tape out is this year.
Mispology's FPGA-based Zebra AI inference accelerator outperformed Nvidia A100, V100, Tesla T4, AWS Inferencia, Google TPUv3, and others, on the ResNet 50 architecture in ML inference efficiency.
The new P4d instances will house eight Nvidia A100 GPUs and 96 Cascade Lake Xeon vCPUs and offer 2.5x times the deep learning performance compared to the last generation, lowering costs up to 60%.