Tsinghua's ACCEL chip: Next-level speed for AI vision
Tsinghua University has unveiled ACCEL, a new photonic chip that could change the game for AI tasks like image recognition and self-driving tech.
This all-analog chip runs at a wild 4.6 peta-OPS—making it hundreds to thousands times faster than current optical or analog chips.
Why is ACCEL such a big deal?
ACCEL isn't just fast—it's insanely efficient.
Built using older, more accessible semiconductor processes, it skips traditional analog-to-digital conversion and uses light (photons) for direct processing.
That means ultra-low latency (just 72 nanoseconds!) and energy use that the team says is over four million times lower than existing high-performance optical and electronic computing architectures, including GPUs and TPUs.
Real-world performance
On real tests, ACCEL scored high: 85.5% accuracy on Fashion-MNIST, over 82% on ImageNet (three-class), and nearly 93% for video recognition—even in low light.
Its super-efficient design could be huge for everything from autonomous cars to next-gen cameras.
How does it work?
Instead of crunching numbers one by one, ACCEL processes tons of data in parallel using photons—basically letting light do the heavy lifting.
It also uses adaptive training methods and a reconfigurable electronic module to correct errors and improve signal quality, especially when things get dark.
In short: this chip could make future AI both smarter and way more energy-friendly.