8.2 A Versatile 7nm Adaptive Compute Acceleration Platform Processor

摘要

As benefits from Moore's Law diminish [1], general-purpose compute platforms like CPUs and GPUs continue to become increasingly power inefficient. The majority of workloads for emerging applications on the edge, like autonomous driving, sensor fusion, robotics, IoT, require flexible high-bandwidth I/O with matched compute and storage. Enterprise workload acceleration in the datacentres requires a heterogenous compute platform that provides the flexibility to be reconfigured in the field (e.g. to implement future neural-network innovations without hardware upgrade). Wireless 5G radios also require a similar mix of flexible I/O matched with high-performance math acceleration for efficient implementation of signal processors for MIMO systems. ACAP is a versatile new platform that addresses several of these domains with >130 INT8 TOPS for power-efficient neural-network implementations, and >100GSPS for CMPLX16 signal processing for MIMO applications.