Photo of Song Yao

Computer & electronics hardware

Song Yao

Challenging GPU’s position in the world of AI

Year Honored


Song Yao, co-founder of DeePhi Tech, who just turned 24 in 2017, was listed in the TR35 China as an entrepreneur—the youngest award winner of the year.

Yao gave up the opportunity to pursue a PhD degree at Carnegie Mellon University, and chose to build a startup because he saw new opportunities ahead. He thinks that in the past, China has always been trying to catch up with their foreign competitors in almost every industry, but now, things are different in the field of artificial intelligence. It’s a rare chance that China is at the same stage with the rest of the world in this cutting-edge technology, and it even has the possibility of leading the game.

AI is driven by three major elements: data, algorithms, and computational power. By accumulating large amounts of data, and with the support of high-performance chips such as GPUs, deep learning can perform valuable data mining, and exceed the accuracy of human recognition, thereby partially realizing the commercial application of deep learning.

Since 2013, Yao’s team has realized that traditional computing platforms such as CPUs and GPUs are not sufficiently efficient or stable, and are also confined by their high cost for practical applications. Therefore, Yao and his research team designed a more efficient framework for neural networks, and brought up the concept of ‘deep compression’—a software-hardware co-design which would speedup the computing with sparse neural networks.

Their products have demonstrated advanced techniques and great potential for future innovation, and attracted the attention of investors and semiconductor manufacturers. DeePhi Tech’s important strategic investors include: Xilinx, the world’s largest FPGA supplier, Mediatek, the global fabless semiconductor company, and Samsung.

Yao believes that the key innovation of DeePhi Tech lies in its combination of hardware acceleration with neural network compression so as to apply deep learning to embedded devices and equip it with programmable architectures. He is convinced that such innovations will bring great breakthroughs to the future of AI application.