News
The energy efficiency of 65nm AI chips far exceeds that of A100! SK Hynix is going big this time
2 min read
Source: news.mydrivers.com
Kuai Technology reported on July 11 that SK Hynix, TetraMem and the University of Southern California jointly made a big deal and released an in-memory computing (IMC) SoC (system on a chip) chip based on memristor (memristor), specializing in edge AI devices. The core selling point of this chip is energy efficiency, which reaches 21.3 TOPS/W at 100MHz and 11.9 TOPS/W at 400MHz. The paper claims that the energy efficiency is an order of magnitude higher than the INT8 mode of NVIDIA A100. But the performance is a bit lagging, with a peak value of only 2.54 TOPS, which is 16 times lower than the requirements of Microsoft Copilot+. The chip is manufactured using the old 65nm process and contains 10 NPUs (Neural Network Processing Units), 9 of which run traditional operations, and 1 of which is specially optimized for depth separable convolution (DWC). The optimized NPU uses a zigzag cross array topology, changing the selection lines to oblique routing, achieving 28 independent 3×3 convolutions in parallel, and maximizing array utilization. The memristor programming accuracy is only about 2 bits. The team used twin array compensation technology to increase it to about 4 bits. The idea is similar to NVIDIA NVFP4, but one uses digital compensation and the other uses analog compensation. In the actual demonstration, MobileNetV1Small was used to run the test, and the end-to-end inference accuracy was 80.36%, which is equivalent to the 4-bit software model. However, only 6 of the 10 NPUs were used, and the paper did not disclose the full real throughput. SK Hynix is responsible for the development and manufacturing of memristor devices, using back-end processes to integrate resistive switching memory cells on top of 65nm CMOS circuits. This paper is essentially a proof of concept. The energy efficiency data is eye-catching but the performance numbers are not impressive. However, for edge AI scenarios that are sensitive to power consumption, this direction deserves attention.