BAIDU DEBUTS FIRST AI ACCELERATOR: Kunlun K200 Is Co-designed for Deep Learning

摘要

Joining other cloud-service providers (CSPs), Baidu is taking a shot at AI mastery by offering the Kunlun K200 SoC. The Chinese company designed the 256-TOPS accelerator for its internal deep-learning workloads. The K200 accelerates common neural-network and SQL operations. On AI inference benchmarks, it matches the power efficiency of Nvidia's T4 card. At the recent Hot Chips conference, Baidu revealed that the Kunlun design features eight clusters combining in-house CPU cores and programmable neural-network engines. The chip's 2.5D packaging supports two stacks of 8GB High Bandwidth Memory (HBM), providing 256GB/s of bandwidth each. Built in Samsung's 14nm technology, the K200 has a moderate power envelope of 150W and connects to the host processor through a x8 PCIe Gen4 interface.