Alibaba Unveils Hanguang 800: The World's Fastest AI Chip Shattering Benchmarks

Alibaba's first self‑developed chip, the Hanguang 800, was officially launched at the Hangzhou Cloud Expo on September 25, where Damo Academy director Zhang Jianfeng showcased what he called the world's strongest AI chip.

In the industry‑standard ResNet‑50 benchmark, the Hanguang 800 achieved an inference throughput of 78,563 IPS, four times higher than the current best AI chip, and an energy‑efficiency ratio of 500 IPS/W, 3.3× higher than the runner‑up.

Zhang said, "In the global chip arena, Alibaba is a newcomer; XuanTie and Hanguang 800 are the first steps of the Flat‑Head (PingTouGe) long march, and we still have a long way to go."

The chip's computing power is equivalent to ten GPUs in Alibaba's City Brain business test; a single Hanguang 800 can replace ten traditional GPUs.

Performance breakthroughs stem from hardware‑software co‑innovation: the hardware uses a self‑designed chip architecture with inference acceleration, while the software integrates Damo Academy's advanced algorithms, deeply optimising CNN and vision workloads, increasing storage density, and enabling large models to run on a single NPU.

Hanguang 800 is already deployed in Alibaba's core internal services. In the City Brain scenario, processing traffic video for Hangzhou's main urban area previously required 40 GPUs with 300 ms latency; with Hanguang 800 only four chips are needed, cutting latency to 150 ms. For the daily influx of one billion product images, traditional GPUs took about one hour for recognition, whereas Hanguang 800 reduces the time to five minutes.

The chip will be offered externally through Alibaba Cloud AI compute services, which launched the same day and claim a 100% improvement in cost‑performance over traditional GPU compute.

In the past six months, PingTouGe released the XuanTie 910 and the WuJian SoC platform. With the launch of Hanguang 800, PingTouGe's end‑cloud integrated full‑stack product line is taking shape, covering processor IP, a one‑stop chip design platform, and AI chips, achieving full coverage of the chip design chain.