Chinese technology giant Alibaba Group has unveiled its first in-house accelerator chip for artificial intelligence (AI) workloads, the Hanguang 800 Neural Processing Unit (NPU).
The growing popularity of deep-learning workloads, which are being used for everything from real-time translation and medical imagery analysis to upscaling pixel art from classic games, has triggered an unusual arms race. Where CPU workloads see Intel and AMD trading blows, and general-purpose GPU (GPGPU) workloads are the battleground for AMD and Nvidia, the need to run training and inference workloads as quickly as possible has led to an explosion of custom chips from companies not previously known for their hardware efforts. Google has its Tensor Processing Unit family, recently launched in dev-friendly form, and now Alibaba has announced the Hanguang 800 Neural Processing Unit (NPU).
Unveiled at the Aspara Computing Conference this week, Alibaba's Hanguang 800 starts its life - as with Google's TPU range - as an internal product designed to accelerate Alibaba's own cloud computing platform. The company has, however, indicated that it plans to mass-produce the part with a view to providing third-party access to its cloud computing customers - but has thus far not announced any plans to sell the hardware itself.
'The launch of Hanguang 800 is an important step in our pursuit of next-generation technologies, boosting computing capabilities that will drive both our current and emerging businesses while improving energy-efficiency,' claims Jeff Zhang, Alibaba Group chief technical officer and president of the company's Cloud Intelligence division. 'In the near future, we plan to empower our clients by providing access through our cloud business to the advanced computing that is made possible by the chip, anytime and anywhere.'
While full performance figures have not been released, Alibaba is claiming some impressive gains from moving to the in-house part: Categorisation and search analysis of the one billion product images uploaded to the company's e-commerce division Taobao each day took the company's platform an hour; with the Hanguang 800 co-processor, Alibaba claims this has been reduced to five minutes.
Hanguang 800 isn't Alibaba's first shot at custom silicon, however: The company's T-Head division, known in Chinese as 'Pintouge' for 'honey-badger,' unveiled a high-performance 16-core RISC-V processor dubbed the CoreXuanTie910 back in July.
October 14 2021 | 15:04