3 ms·

Everyone knows that a good reimplementation of CUDA will make them trillions but despite 15 years, no one has made anything even close to it. There’s profit to

Everyone knows that a good reimplementation of CUDA will make them trillions but despite 15 years, no one has made anything even close to it. There’s profit to be made and no entrants. How can we explain this? If your answer is “AMD et al tried” you aren’t paying attention to just how pathetic and shallow their efforts have been and still are.

cpgxiii 2y ago
Every GPU vendor knows that actually supporting a reimplementation of CUDA for their hardware immediately signs them up for two problems: 1. Protracted lawsuits with the (now) second-most-valuable company in the world. 2. Any significant customers being loudly reminded by Nvidia that the EULA for CUDA tools prohibits their use on non-Nvidia hardware. The problem is not the technical challenge of reimplementing the CUDA API or tools. The problem is the users want CUDA, not something that looks and behaves like CUDA but requires them to load a different set of libraries.
- jmalicki 2y ago
  These days, you don't really need to support CUDA, just PyTorch gets you most of the way there. You have Google TPUs, Amazon Inferon, a ton of smaller players that aren't quite getting traction - but there's a ton of investment there.
creer 2y ago
In addition to the other response (risks), the economy is not a stable homogeneous field. None of the economy is a stable homogeneous field. Instead it's a bunch of discrete individuals living their life and each focusing on just a few things. And this goes both ways. For a competitor to form requires one small team of founders who are insane enough (in a good way) to believe they should try it, and a funding source that's ready to take the gamble on/with them. So you might get a fantastically successful company "out of nowhere" (Amazon, say). Or you might get a field where nobody is at the moment on hand to try it. Or several do try it and fail silently and you never hear of them. They are individuals, not some mathematical process. See also, even "efficient market" does not refer to THAT.

cpgxiii 2y ago

Every GPU vendor knows that actually supporting a reimplementation of CUDA for their hardware immediately signs them up for two problems: 1. Protracted lawsuits with the (now) second-most-valuable company in the world. 2. Any significant customers being loudly reminded by Nvidia that the EULA for CUDA tools prohibits their use on non-Nvidia hardware. The problem is not the technical challenge of reimplementing the CUDA API or tools. The problem is the users want CUDA, not something that looks and behaves like CUDA but requires them to load a different set of libraries.

jmalicki 2y ago
These days, you don't really need to support CUDA, just PyTorch gets you most of the way there. You have Google TPUs, Amazon Inferon, a ton of smaller players that aren't quite getting traction - but there's a ton of investment there.

creer 2y ago

In addition to the other response (risks), the economy is not a stable homogeneous field. None of the economy is a stable homogeneous field. Instead it's a bunch of discrete individuals living their life and each focusing on just a few things. And this goes both ways. For a competitor to form requires one small team of founders who are insane enough (in a good way) to believe they should try it, and a funding source that's ready to take the gamble on/with them. So you might get a fantastically successful company "out of nowhere" (Amazon, say). Or you might get a field where nobody is at the moment on hand to try it. Or several do try it and fail silently and you never hear of them. They are individuals, not some mathematical process. See also, even "efficient market" does not refer to THAT.