This assumes the chips don’t get exponentially better.
GPUs are actually bad for AI. TPUs too. They’re available and general purpose but what you really want is something like a specialized tensor FPGA that can be loaded with the model. Memory and compute colocated.
There is research on this and a huge market but it takes a long time to go from design to prototype to production for something really novel. It’s also very expensive.
A lot of people are afraid AI is a big bubble. The more it becomes apparent that financial bubbles or not the tech is here to stay the more people will be comfortable making long term investments in building chips that are really optimal for it.
Even general chips get more efficient. Ten years ago my laptop was a lap roaster when I ran a big build on it. Now it’s barely warm and it’s twice as fast with four times the RAM.
Of course if there’s infinite demand for AI that won’t matter, but there isn’t infinite demand for anything.
GPUs are actually bad for AI. TPUs too. They’re available and general purpose but what you really want is something like a specialized tensor FPGA that can be loaded with the model. Memory and compute colocated.
There is research on this and a huge market but it takes a long time to go from design to prototype to production for something really novel. It’s also very expensive.
A lot of people are afraid AI is a big bubble. The more it becomes apparent that financial bubbles or not the tech is here to stay the more people will be comfortable making long term investments in building chips that are really optimal for it.
Even general chips get more efficient. Ten years ago my laptop was a lap roaster when I ran a big build on it. Now it’s barely warm and it’s twice as fast with four times the RAM.
Of course if there’s infinite demand for AI that won’t matter, but there isn’t infinite demand for anything.