OpenAI is reportedly developing its own model-specific hardware.
1. It’s a small model. Nowhere near the big leading ones.
2. They built custom chips and trained the model to work with it. They explain it somewhere on the website. I believe they are working on a v2 now with a bigger smarter model.
I believe other companies including OpenAI are working on similar projects with custom chips.
@NickADobos wtf....how?
No Digg Deeper questions have been answered for this story yet.
@NickADobos So interesting!! Do you think building custom chips for each model is the future?
@MatthewBerman Also IIRC they actually burn it into the chip. So once you train the model and build the chips you can’t upgrade it or swap to a new model. You would need to build brand new chips.
So it’s a very different technique vs a chip that can run any type of inference
@MatthewBerman Also IIRC they actually burn it into the chip. So once you train the model and build the chips you can’t upgrade it or swap to a new model. You would need to build brand new chips.
So it’s a very different technique vs a chip that can run any type of inference
1. It’s a small model. Nowhere near the big leading ones.
2. They built custom chips and trained the model to work with it. They explain it somewhere on the website. I believe they are working on a v2 now with a bigger smarter model.
I believe other companies including OpenAI are working on similar projects with custom chips.
@MatthewBerman Probably not?
Hard to do because models are changing so quickly.
Also knowledge decays rapidly. So out of date models perform worse over time.
But for workloads that require speed and maybe don’t need as much intellect, well this clearly works!
@NickADobos So interesting!! Do you think building custom chips for each model is the future?