Alibaba’s DAMO Academy has teased a chip that stacks logic and memory in 3D, and Chinese press suggest it represents an architecture that can bust the Von Neumann bottleneck.
The DAMO Academy is Alibaba’s blue sky research outfit. Founder Jack Ma dedicated it to “Discovery, Adventure, Momentum and Outlook” at its 2017 launch.
Both reports suggest the chip could bust the von Neumann bottleneck – the limitation encountered by most computer systems because internal bus speeds mean CPUs must wait for data to be fetched from memory. Tricks like caches, parallel computing, and speculative execution work around the bottleneck, as do recent innovations like Apple’s M1 architecture and other system-on-chips that place processor and memory on a single die and therefore remove the need to communicate with memory outside the processor. But the bottleneck is considered a fundamental problem in computers that employ the von Neumann architecture of discrete processing and storage.
Alibaba has indicated to The Register that the chip exists, but offered no further details at the time of writing.
The Register has also learned that a presentation at February 2022’s International Solid-State Circuits Conference – an important event for chip design boffins – features a talk tantalisingly entitled “184QPS/W 64Mb/mm2 3D Logic-to-DRAM Hybrid Bonding with Process-Near-Memory Engine for Recommendation System” in its stream on ML Chips for Emerging Applications. Of the 19 co-authors listed for the talk, 15 work at the DAMO Academy.
While we wait for that talk – and whether it delivers on the hype Chinese media have applied to the 3D chip – Alibaba has given us all something we can inspect: the source code for its Yun on Chip (YoC), full-stack technology development platform for IoT devices.
YoC uses the RISC-V architecture and its SIMD acceleration instructions, and was co-developed with Hangzhou C-SKY Microsystem in 2015. It can be used for various applications including Bluetooth, Wi-Fi, audio, and multimedia. Alibaba told The Register it can also be tuned to the needs of AI acceleration, audio processing, mesh networking and motor control.
The Chinese colossus hasn’t said where the Yun designs will appear, but the Xuantie cores it open sourced in October 2021 popped up on the GitHub repo of its chip design subsidiary T-Head Semiconductor. ®