China: World’s 1st light-based AI chip beats NVIDIA H100 in energy efficiency
Ateam of scientists from Beijing has announced a groundbreaking advancement in artificial intelligence (AI) technology with the development of the world’s first fully optical AI chip.
This innovative chip, known as Taichi-II, represents a significant leap forward in both efficiency and performance, surpassing even the renowned NVIDIA Corp. NVDA H100 GPU in energy efficiency.
The research team, led by professors Fang Lu and Dai Qionghai from Tsinghua University, unveiled their findings on Wednesday.
A leap beyond: The Taichi-II chip’s superiority
The Taichi-II chip represents a substantial advancement from its predecessor, the Taichi chip, which had already set impressive records. Earlier this year, the researchers announced that the original Taichi chip had exceeded the energy efficiency of NVIDIA’s H100 GPU by over a thousand times, as
reported by
South Morning China Post (
SCMP).
Now, the Taichi-II chip has further elevated this benchmark, showcasing superior performance across various scenarios.
The study led by Professors Fang Lu and Dai Qionghai highlights Taichi-II’s capability to transform
AI training and modeling. Unlike traditional methods that rely on electronic computers for training, the Taichi-II leverages optical processes, making it more efficient and significantly enhancing performance.
In practical terms, the Taichi-II chip has demonstrated remarkable advancements in several areas. It has expedited the training of optical networks containing millions of parameters by an order of magnitude and improved the accuracy of classification tasks by 40 percent.
In complex imaging scenarios, its energy efficiency in low-light conditions has improved by six orders of magnitude.
Innovative approach: FFM learning
The
development of the Taichi-II chip is marked by its use of a novel approach called fully forward mode (FFM) learning. This technique allows for a computer-intensive training process to be conducted directly on the optical chip, enabling parallel processing of machine learning tasks.
Xue Zhiwei, lead author of the study and a doctoral student, emphasized that this architecture supports high-precision training and is well-suited for large-scale network training.
“Our research envisions a future where these chips form the foundation of optical computing power for AI model construction,” Fang Lu stated.
The FFM learning method capitalizes on high-speed optical modulators and detectors, which could potentially outperform GPUs in accelerated learning scenarios. This innovation opens new possibilities for optical computing, moving it from theoretical concepts to practical, large-scale applications.
Implications and future prospects
The timing of Taichi-II’s debut is particularly notable. As the US has imposed restrictions on China’s access to advanced GPUs for
AI training, the Taichi-II chip offers a viable alternative that could help mitigate these limitations.
Additionally, the performance of Taichi-II comes amid reports that NVIDIA’s high-tech AI chips may be making their way into the hands of Chinese military officials, potentially influencing China’s technological advancements.
Taichi-II demonstrates up to a thousandfold improvement in energy efficiency and accelerates AI training with cutting-edge optical processes.
interestingengineering.com
Fully forward mode training for optical neural networks
We present fully forward mode learning, which conducts machine learning operations on site, leading to faster learning and promoting advancement in numerous fields.
www.nature.com
www.computerworld.com
