Comment from SA concerning ASIC versus GPU
"how about the data center customers using their own chips?"
How's that working out for google? Arguably they understand the AI problem better than anyone, they have the most insight into their own hardware, software, deployment, uses cases and how effective a home grown solution can be. And yet they continue to buy a ton of GPUs. And that's after 4 revisions of their own TPU.
The difference between a purpose built ASIC and a general purpose GPU is programmability and flexibility which provides utility over a wide range of use cases. Maybe someone will break it open at some point, but we've been hearing that threat for 6 or 7 years now.
The problem the asic guys have now, as Graphcore and Cerebras have discovered, is the performance of the entire data center, storage, networking, memory, as well as compute, are critical to making large gains. The effort has moved from a creating a novel circuit to rearchitecting these giant machines. |
