SK Hynix and partners build new chip to make edge AI more efficient

SK Hynix, along with TetraMem and the University of Southern California, has developed a new type of chip that uses memristors to handle both memory and computing tasks. This chip is designed to make AI on small devices like phones and sensors more efficient and use less power.

The chip uses a RISC-V processor to manage tasks and has 10 special units called NPUs. Together, they can reach a computing power of about 2.54 TOPS. By using a clever design with two memory arrays, the chip achieves a weight precision of about 4 bits and delivers 80.36% accuracy in tests, matching what software can do.

At a low speed of 100 MHz, each NPU uses very little power while still being fast, reaching 21.3 TOPS per watt. This technology could help make future smart devices, like cameras and wearables, run AI tasks without draining the battery.