IBM Research has unveiled a mixed-signal analog Chip designed for AI inferencing that can match the performance of digital counterparts like GPUs but with significantly lower power consumption. The chip, part of IBM’s HERMES project, consists of 64 tiles or compute cores, making it more powerful than the previous 34-tile version. The chip utilizes a combination of digital circuits and phase-change memory to perform matrix-vector multiplications on network weights stored on the chip. It also features on-chip communication networks and functions necessary for processing convolutional layers.

While most current AI architectures have separate memory and processing units that require constant data shuffling, IBM’s chip follows an approach called analog in-memory computing (AIMC). It uses phase-change memory cells to store weights as analog values and perform computations. Each core of the chip contains a PCM crossbar array capable of storing a 256×256 weight matrix and performing analog matrix-vector multiplication.

The chip achieves near-software-equivalent inference accuracy, with a reported accuracy of 92.81% on the CIFAR-10 image dataset. It also boasts a high multiplication throughput and energy efficiency, performing 400 giga-operations per second per square millimeter (400 GOPS/mm2), which is more than 15 times higher than previous multicore chips based on resistive memory.

IBM’s research paper suggests that with additional digital circuitry, chips like this could enable fully pipelined end-to-end inference workloads. However, further improvements in weight density are needed for AIMC accelerators to compete with existing digital solutions like GPUs.

Last year, another research paper published in Nature described an experimental chip that stored weights in resistive RAM (RRAM), consuming less than 2 microwatts of power for real-time keyword spotting. This stands in contrast to the increasing power requirements of GPUs used for AI, with some data centers supporting up to 70 kilowatts per rack for AI infrastructure.

The post IBM Develops Low-Power Analog Chip for AI Inferencing appeared first on TS2 SPACE.