Cerebras Unveils Wafer Scale Engine Two (WSE2): 2.6T Transistors, 100% Yield
anandtech.com
anandtech.com
The architectures we have are shaped by the hardware they run on, I think more than people realize. Different hardware will need different architectures to be fully utilized. Unfortunately there are so few of these chips in existence that it will be hard to get enough people using them to really discover how to best use them.
I wish we heard more about what people are doing with the first version. From the article it sounds more like fluid dynamics simulation than the AI stuff Cerebras likes to talk about. Wide availability of GPUs really kicked off the deep learning revolution, and it would be tough for this chip to have the same kind of impact with only "dozens" of customers worldwide. But those rare customers could potentially do some really, really cool stuff.