The ACCEL chip was tested against various image recognition tasks and showed high accuracy levels while demonstrating superior system robustness in low-light conditions. The chip's architecture allows it to reconfigure analog pathways to accelerate specific tasks, making it an analog version of an Application-Specific Integrated Circuit (ASIC). Despite being manufactured on a standard 180-nm CMOS technology, further performance and efficiency improvements could be gained from miniaturizing the process towards lower CMOS nodes. The main challenge for implementing such chips at scale is the current low manufacturing throughput and industry adaptation.
Key takeaways:
- A new AI processing chip called ACCEL, developed by Tsinghua University in China, is capable of delivering over 3,000 times the performance of an Nvidia A100 with an energy consumption that’s four million times lower.
- The ACCEL chip uses photonic and analog computing in a specialized architecture, with 99% of its operation implemented within the optical system. This allows for fewer energy requirements and less waste heat compared to digital systems.
- The chip has been tested on various vision tasks with high accuracy levels and has shown superior system robustness in low-light conditions.
- Despite the impressive performance of the ACCEL chip, its manufacturing is currently too low to serve anything other than research efforts and prototypical work. However, the potential for this technology in the market is significant.