ML//TPU

The TPU is Google's custom ASIC for ML training and inference at datacenter scale, exposed as cloud instances (Google Cloud TPU). It represents the extreme end of ML hardware specialization: math embedded as silicon.

The TPU is Google's custom ASIC for ML training and inference at datacenter scale, exposed as cloud instances (Google Cloud TPU). It represents the extreme end of ML hardware specialization: math embedded as silicon.

It uses a systolic array architecture to stream matrix operations with minimal memory access.

TPU v4 pods interconnect thousands of chips through custom high-bandwidth networks.

Both the TPU and the NPU are ASICs optimized for tensor multiplication.