Cambricon 690
#8 v AI inferenční hardwareCambricon · v690 · 2× · naposledy 11. 9. 2026
The Cambricon 690 (also called Siyuan/Sīyuán 690 or MLU690) is the upcoming AI accelerator chip from Cambricon Technologies in Beijing, expected to match Nvidia's H100 in performance. It is being developed as the successor to the Siyuan 590, which was introduced in 2023, and will be manufactured on the SMIC N+2 process (7nm-class). According to Reuters reports from September 2026, the chip had not been formally launched at the time of research but was already in testing phases or, according to some sources, in mass production. No official price has been communicated, though price increases of 20-30% compared to previous internal indications have been reported. An official product page with complete technical specifications is not yet publicly available.
Vlastnosti
| Manufacturing Process (nm) | SMIC N+2 process, 7nm-class (DUV lithography) |
| Platform | OAM (Open Accelerator Module) form factor, server/cloud accelerator for training and inference |
| Price | No official price disclosed; Reuters reports Cambricon repriced internal indications by 20-30% versus levels indicated two months earlier |
| Compute Performance (FLOPS/TOPS) | Per industry report >700 TFLOPS FP16 (dual-die design); per expert interview FP16 benchmark score of 813 |
| Release Date | Not formally launched yet (as of September 10, 2026); TrendForce suggests mass production may slip to H2 2026 |
| Memory | 196GB HBM3 memory (per expert interview with ByteDance specialist / industry report) |
| Availability | Not formally launched yet; planned 2026 volume: up to 300,000 units (Siyuan 590 + 690 combined) |