By: RichardC (tich.delete@this.pobox.com), May 14, 2013 5:27 pm
Room: Moderated Discussions
Ricardo B (ricardo.b.delete@this.xxxxx.xx) on May 14, 2013 3:00 pm wrote:
> We've given you a lot of examples of common use cases for multi-core and SMT on client computers.
No. You've given a bunch of examples, most of which a) aren't frequently used on
desktop/laptop, and/or b) are done with specialized hardware on modern cpu's
(encryption, video transcode), and/or c) don't go noticeably faster on 4C/8T than
on 4C/4T (many - perhaps most - current games, some flavors of compression).
Quad-cores are good. Quad-cores with SMT ? Not noticeably better for most people
[Yeah, yeah, so if you're doing video editing or compiling 1M-line programs,
maybe it helps. But 99% of people aren't].
> The argument is that, because of these users, in a weird world where SMT was
> not available, Intel might have gone for more thinner cores (like AMD).
It's still a really bizarre argument, that Intel would have chosen to ignore the
predominant case of single-threaded apps, in favor of optimizing for the small
number of highly scalable apps.
> We can also give you examples for single thread performance on servers. All in all, Amdhal's
> law applies and in many tasks lots of slow cores can't keep up with a few fast ones.
Absolutely. But there are a couple of really important differences about servers:
a) they're usually serving many clients, and thus have some degree of inherent
parallelism in parts of the workload, and b) they're spending most of their time in a
small number of highly-tuned (and often very expensive) programs, rather than having
hundreds of different programs. So the necessary - very large and expensive - effort
is applied to that server software to make it scale quite well on various hardware
configurations.
> We've given you a lot of examples of common use cases for multi-core and SMT on client computers.
No. You've given a bunch of examples, most of which a) aren't frequently used on
desktop/laptop, and/or b) are done with specialized hardware on modern cpu's
(encryption, video transcode), and/or c) don't go noticeably faster on 4C/8T than
on 4C/4T (many - perhaps most - current games, some flavors of compression).
Quad-cores are good. Quad-cores with SMT ? Not noticeably better for most people
[Yeah, yeah, so if you're doing video editing or compiling 1M-line programs,
maybe it helps. But 99% of people aren't].
> The argument is that, because of these users, in a weird world where SMT was
> not available, Intel might have gone for more thinner cores (like AMD).
It's still a really bizarre argument, that Intel would have chosen to ignore the
predominant case of single-threaded apps, in favor of optimizing for the small
number of highly scalable apps.
> We can also give you examples for single thread performance on servers. All in all, Amdhal's
> law applies and in many tasks lots of slow cores can't keep up with a few fast ones.
Absolutely. But there are a couple of really important differences about servers:
a) they're usually serving many clients, and thus have some degree of inherent
parallelism in parts of the workload, and b) they're spending most of their time in a
small number of highly-tuned (and often very expensive) programs, rather than having
hundreds of different programs. So the necessary - very large and expensive - effort
is applied to that server software to make it scale quite well on various hardware
configurations.
