[mvanveen]

In many cases within regulated environments getting access to GPU in e.g. a customer's on-prem solution is prohibitive.

In enterprise ML contexts on tabular data problems we found there are a lot of cases where even training can be greatly sped up by leveraging AVX instruction support in e.g. tensorflow builds. The gains from AVX instructions could boost training time by ~20% on the GAN use cases I profiled.

[btilly]

So you're saying that we need to solve organizational stupidity by complicating all chips for everyone so that bad organizations can get a performance boost on specialized tasks?

What percentage of consumers with these chips installed do you think are getting a performance win? Do you think it might be as high as 1%?

Do you think that the same resources devoted elsewhere might be worth more than 1% to that 99%? Whether that is in reduced cost, reduced bugs, or a boost for more widely used operations.

Yes, if you target a specific use case for a specific set of people, you can give them a nice win. But you shouldn't lose sight of the fact that CPUs cover a lot of use cases for a lot of people. And simplifying then focusing on the core mission is better for everyone in the end.

[anoncareer0212]

Linus somewhat waved away the idea of tradeoffs, which is fine, he was speaking in generalities.

Turning factual reports of where the tradeoff was helpful into a strawman insulting the reporter, and the users who benefit, is neither charitable nor illuminating.