China currently faces a blockade preventing access to the most advanced manufacturing processes, a situation that might persist for a considerable time. Nevertheless, it’s possible that Chinese engineers have devised a method to make their chips, manufactured using older process nodes, perform as rapidly as Western chips built on state-of-the-art nodes.
Wei Shaojun, the Vice Chairman of the China Semiconductor Industry Association, has announced the development of an AI processor fabricated using 14nm technology that purportedly matches the performance of Nvidia’s latest 4nm chips.
Complete specifics remain scarce, as this pertains to an experimental unit rather than a mass-market product. However, reports suggest that the breakthrough lies in integrating the processor and DRAM memory via a novel 3D packaging technique.
If these claims hold true, this technology allows the chip, despite its outdated fabrication process, to achieve an energy efficiency of 2 TFLOPS per watt, boasting a total throughput of 120 TFLOPS while consuming only 60 W. In terms of power efficiency, this surpasses Nvidia’s current offerings, though its peak raw performance trails. However, this scalability deficit can be addressed by increasing the sheer number of processors, and the superior energy efficiency permits such scaling without unduly increasing overall power draw.
Unfortunately, there is no clarity yet on when we can anticipate seeing the first commercial chips incorporating this technology on the market.
About the Author
