A recent research paper from Tsinghua University in China introduces an innovative AI processing chip named ACCEL, designed specifically for computer vision tasks. This chip utilizes a combination of electronic and light computing, leveraging photonic and analog computing techniques to deliver exceptional performance. The paper, published in Nature, highlights that the simulated ACCEL processor achieves 4,600 tera-operations per second (TOPS) in vision tasks, providing a substantial 3.7X performance advantage over Nvidia’s A100 in image classification workloads.

Researchers working on an all-analog photoelectronic chip gather for a group photo at Tsinghua University in Beijing, China on April 20, 2021. (Image courtesy: Tsinghua University/Handout via Xinhua)

The paper emphasizes the significance of accuracy in computer vision tasks and demonstrates the chip's effectiveness through competitive levels of accuracy in various tasks. ACCEL's architecture operates predominantly through diffractive optical analog computing (OAC) assisted by electronic analog computing (EAC), with 99% of its operation implemented within the optical system. This approach aims to address constraints found in other vision architectures while significantly improving energy efficiency compared to non-analog approaches.

The use of photonic, optical systems in ACCEL reduces energy requirements, minimizes electron waste in thermal dissipation, and enables higher operating speeds limited only by light itself. Additionally, the chip showcases low computing latency and impressive frame generation throughput, indicating its potential to significantly enhance deep learning processing in computing-vision tasks.

ACCEL's design resembles that of an Application-Specific Integrated Circuit (ASIC) and features an electronic analog computing (EAC) unit capable of reconfiguring analog pathways to accelerate specific tasks. The research team emphasizes the practical implementation of this new architecture to address major national and public needs.

ACCEL's manufacturing at a 180-nm CMOS technology for the Electronic Analog Computing unit (EAC) suggests potential further efficiency improvements through miniaturization. However, large-scale deployment of high-performance AI analog chips like ACCEL currently faces challenges related to manufacturing throughput and industry adaptation.

The introduction of ACCEL signifies a major milestone in AI chip development, and while challenges remain for its widespread deployment, the experimental results showcase its potential for transforming AI computing.

Link to the full document https://www.nature.com/articles/s41586-023-06558-8?utm_medium=affiliate&utm_source=commission_junction&utm_campaign=CONR_PF018_ECOM_GL_PHSS_ALWYS_DEEPLINK&utm_content=textlink&utm_term=PID100046186&CJEVENT=abb5a900825011ee804902600a18ba73