"For LLMs, IBM's NorthPole chip overcomes the tradeoff between speed and efficiency."
IBM's NorthPole chips have a different architecture from GPUs, more directly inspired by the brain.
This page has funky graph that has energy efficiency (not energy) on the vertical axis and latency on the horizontal axis -- but in reverse order, so the slowest latency is on the right. Both axes are logarithmic. The only reason I can think of why it's done this way is to make "better" up on the vertical axis and to the right on the horizontal axis. Better energy efficiency is good so you want to go up on the vertical axis. Low latency is good so you want to go to the right on the horizontal axis. With this set up they put the NorthStar chip in the upper right corner.
I wonder if there's a possibility of this competing commercially with Nvidia.
For LLMs, IBM's NorthPole chip overcomes the tradeoff between speed and efficiency
There are no comments yet.