China's new AI chip outperforms NVIDIA's most powerful GPU
A team of researchers from Beijing, led by Professors Fang Lu and Dai Qionghai of Tsinghua University, has unveiled the world's first fully optical artificial intelligence (AI) chip. Named Taichi-II, this groundbreaking innovation marks a significant milestone in the field of optical computing. The chip has outperformed NVIDIA's H100 GPU in terms of energy efficiency and performance.
It offers superior performance in AI training
The Taichi-II has demonstrated exceptional performance across various scenarios than its predecessor, the Taichi chip. It has significantly improved AI training and modeling by using optical processes instead of traditional electronic computers. This approach has expedited the training of optical networks containing millions of parameters, and boosted the accuracy of classification tasks by about 40%.
Innovative approach boosts energy efficiency
In complex imaging scenarios, the Taichi-II chip's energy efficiency has improved by six orders of magnitude in low-light conditions. The chip employs a unique method known as fully forward mode (FFM) learning. This approach allows for computer-intensive training processes to be done directly on the optical chip, permitting parallel processing of machine learning tasks.
FFM learning method: A game-changer in optical computing
The FFM learning method utilizes high-speed optical modulators and detectors that could potentially outpace GPUs in accelerated learning scenarios. This innovation moves optical computing from theoretical concepts to practical and large-scale applications. Xue Zhiwei, the lead author of the study and a doctoral student, highlighted that this architecture supports high-precision training and is well-suited for large-scale network training.
What are the future prospects?
The debut of Taichi-II is particularly significant. With the US restricting China's access to advanced GPUs for AI training, Taichi-II presents a viable alternative to address these limitations.
