Researchers at Peking University recently developed a special analog chip that can process data 1,000 times faster and 100 times more efficiently than Nvidia H100 and AMD Vega 20 GPUs. Instead of digital processing with binary values, it uses analog processing via RRAM cells that store and process data as continuous electrical signals.
In a study published Oct. 13 in the journal Nature Electronics, the team describes how the chip solves two major problems of digital systems. The first is related to high power consumption, and the second is limited data throughput. When tested on complex communication tasks, the chip achieved the accuracy of digital processors, but used 100 times less electricity.
With additional tweaks, the chip outperformed the performance of high-end GPUs by up to 1,000 cards. The key advantage is that data is processed within the chip itself, without the need for transfers between memory and the processor.
Dedicated RRAM cells were arranged in two circuits: one for rapid evaluation and the other for precise refinement of the results. The chip was fabricated using a conventional CMOS process, enabling mass production.
The team plans to develop larger chips for even more demanding tasks in the future.<\/p>\n