IBM is developing an analogue AI chip that might just save our GPUs for gaming
IBM claims to have cooked up a new mixed-signal part-analogue chip using a combination of phase-change memory and digital circuits that it's claiming will match GPU performance when it comes to AI inferencing but do so at much greater efficiency.
No, we don't entirely understand that either. But the implications are easy enough to grasp. If this chip takes off, it could put a cap on the skyrocketing demand for GPUs used in AI processing and save them for, you know, gaming.
According to El Reg, this isn't the first such chip IBM has produced. But it's on a much larger scale and is claimed to demonstrate many of the building blocks that will be needed to deliver a viable low-power analogue AI inference accelerator chip.
One of the main existing bottlenecks for AI inferencing involves shunting data between the memory and processing units, which slows processing and costs power. As IBM explains in a recent paper, its chip does it differently, using phase-change memory (PCM) cells to store inferencing weights as an analogue value and also perform computations.
It's an approach known as analogue in-memory computing and it basically means that you do the compute and memory storage in the same place and so—hey presto—no more data shunting, less power consumption and more performance.
Things get more complex when you start describing the scale and scope of the weighting matrices that the chip can natively support. So, we won't go there for fear of instantly butting up against the exceedingly compact limits of our competence in such matters.
But one thing is fore sure. AI-processing power consumption is getting out of hand. An AI inferencing rack reportedly sucks up approaching 10 times the power of a «normal» server rack. So, a
Read more on pcgamer.com
