For a TPU focused on inference the name of the game is memory bandwidth. How much of the available bandwidth you can extract for as little logic/area/power as you can.
Also can't keep them closed source if you do that.
Also, it's hard to get fab capacity for any project. Let alone something so experimental.
Also: Here is our recursive self-improvement hard at work...
The obvious next step is to get enough memory throughput to run that SOTA model itself so that it develop its own hardware.
But perhaps the more interesting question is this: Can an AI be given a big FPGA and design a model architecture that takes advantage of the fabric being reconfigurable.
But it's complicated for other reasons, one being that the number of parameters for frontier models (especially with MoE models) are so high, and not always utilized (once again, thanks to MoE) that it would actually be incredibly cost prohibitive, if not impossible, to attempt to make giga-chips that would allow running it.
I definitely do believe that we will see more and more specialized chips over time, but putting the entire model on a chip is still a ways away.
I believe Taalas has a heavily handicapped llama 8-billion parameter model. And it still pulls >200W to run.
I can't imagine how anthropic or open ai would be able to burn a multi-trillion parameter model on a chip, we just aren't there yet.
It's why everyone and their dog runs these things on GPUs. When a new model supercedes the previous one, so long as you've got the memory for it your chips aren't obsolete.
I'm looking forward to someone picking a model to be "good enough" (say, qwen 4.0 or something) and selling them as peripheral hardware
So companies try to maximize the memory bandwidth they can get, balancing tradeoffs of power/area/programability of their chip. Right now they feel like the economy on power/area is not worth the decrease in programability/flexibility.
Much of a model are weights, and high-density ROMs are very very very hard.
fsbonetto•55m ago