Siyuan (MLU590/690) by Cambricon: status and timeline

Siyuan (MLU590/690)

Cambricon's Siyuan accelerators using HBM memory and custom compute units for AI training and inference. The Siyuan 590 is in volume production; the 690 is the next-generation variant.

DeployStage 4 of 5

MLU590 in volume production; MLU690 in development.

Checked 10 Oct0 updates this week

Milestones

Next · Siyuan 690 development

Siyuan 290 launched2021Complete.
Siyuan 590 volume production2024Complete.
Siyuan 690 developmentTarget 2026Current milestone.

Upcoming

2026Siyuan 690 development (next)

Current obstacles

Process node disadvantageManufacturing through SMIC limits access to cutting-edge process nodes, affecting performance per watt compared to leading competitors.
Software ecosystem maturityMost AI frameworks prioritize NVIDIA CUDA, requiring significant effort to achieve comparable software support.
Export restrictionsUS export controls limit the addressable market to China and select allied nations.

Physics limits

Process technology constraintsAccess limited to SMIC's process nodes rather than cutting-edge foundries.
Performance scalingPerformance per watt depends on both compute efficiency and memory architecture, with process node limitations affecting overall efficiency.

How it works

3 parts

HBM memory

High-bandwidth memory

HBM provides higher memory bandwidth than GDDR, critical for AI workloads. Cambricon integrates multiple HBM stacks per accelerator.

Custom compute units

Matrix compute engines

Siyuan uses custom compute units optimized for matrix operations in deep learning.

Software

Canopy framework

Cambricon provides tools for porting models from CUDA, though adoption remains limited compared to NVIDIA's ecosystem.

About Cambricon

The team behind Siyuan (MLU590/690)

Cambricon

MLU accelerators, AI chips, Siyuan, China semiconductor

Cambricon is China's leading domestic AI chip maker, designing MLU (Machine Learning Unit) accelerators for cloud servers and edge computing. Founded in 2016, the company went public on Shanghai's STAR in 2020.

China's leading domestic AI chip maker; Siyuan MLU series is primary alternative to NVIDIA within China.
IPO on Shanghai STAR in 2020, raising $369 million.
Designs and manufactures Siyuan series accelerators via SMIC partnership.

cambricon.com (opens cambricon.com)Company profile →

Founded: 201610 yrs
Headquarters: China
Status: Public
Listed: 688256Shanghai STAR
Valuation: public
Last round: IPO on Shanghai STAR board · $369M20 Jul 2020
Works in: ComputingAI accelerators & CPUs
Coverage: 1 program · 0 updateschecked 10 Oct
Customers: Chinese cloud providers
Suppliers: SMIC
People: Chen Tianshi (opens en.wikipedia.org)Founder & CEO