4 ms·
While I love Linus and I think he's excellent at what he does, he's utterly missing the fact that things like AVX-512 sell chips, and that's what Intel's in the
by shadowgovt 6y ago
While I love Linus and I think he's excellent at what he does, he's utterly missing the fact that things like AVX-512 sell chips, and that's what Intel's in the business of doing. And best practice in how one writes software will increasingly look like "write it so it executes well in a massively-parallel context" if the driver for chip sales is massively-parallel problems.
We've smacked pretty hard into the wall of how fast we can make CPUs by miniaturizing components, and the low-hanging fruit now is parallelization, predictive execution, and the whole host of ways to do more per clock cycle, not speed up clock cycles. But that flies in the face of the traditional embarassingly-serial pattern of the x86 instruction set and computing environment. Much of what Intel's doing these days is trying to open up opportunities for people to change tools to speed up code without having to boil the ocean by throwing the whole serial-instruction model completely out the window (even though, increasingly, code written to that instruction set is really operating like a language that is emulated by the underlying parallel-and-predictive CPU hardware).
I think what Linus calls "regular code" isn't going to move chips and we've wrung all the cheap optimizations out of that critical path (and, depending on the details of what is meant by "regular," such code is increasingly going to ram up against an efficiency ceiling unless it can be moved to a model of "Prepare work to do in parallel, execute in parallel, merge results and present to UI").
- colinmhayes 6y agoPretty sure this is exactly what Linus is complaining about. It's not that he doesn't get it, he's upset that intel is spending resources on what he sees as advertising instead of making the product better.
- shadowgovt 6y agoBut more direct access to parallel execution hardware is making the product better. It's not making it better for his use case, which is not surprising because Intel is in the business of selling chips, not optimizing for the Linux kernel use case.
- craftinator 6y agoWow, you've turned the "actually useful" reasoning into "marketing useful" reasoning. I'm speechless.
- netcan 6y agoMarketing isn't the only thing that sells chips. OP is talking about actual usefulness. It's not that useful to Linus and what he needs. OP is arguing that it is useful to others, doing different things, and that those use cases sell chips.
- HelloNurse 6y ago> he's utterly missing the fact that things like AVX-512 sell chips, and that's what Intel's in the business of doing. But wouldn't something more useful than AVX-512 sell more chips? Massively parallel floating point workloads on the CPU are a modest niche compared to other ways Intel could have spent engineering effort and chip area, like not ruining the performance of just about everything else with security vulnerabilities.
- shadowgovt 6y agoNobody saw those security vulnerabilities coming. And ironically, they were introduced due to changes that are intended to make the chips faster without explicit parallelization. This goes back to the problem that a modern architecture for a CPU is a massively parallel machine emulating an instruction set design for an embarrassingly serial machine based on best practices from the 1970s.