Richard Cownie (tich@pobox.com) on 5/10/11 wrote:
---------------------------
>David Kanter (dkanter@realworldtech.com) on 5/10/11 wrote:
>---------------------------
>
>>Color me skeptical. OpenCL isn't that widely used, and I'm not too sure about
>>Grand Central. I agree the pieces are in place...but that's like 5-10 years out.
>
>It doesn't have to be "widely used"; it has to be used in
>the few critical apps that cover a lot of common usage.
>Video transcoding (well, that may be in dedicated hardware);
>photoshop; video editing. And since Apple supplies much of that
>themselves, with iMovie, iTunes, Final Cut Pro etc, they
>can make it happen quickly - it seems that's what they've
>done with Final Cut Pro X already. Not that I really think
>anyone wants to do video editing on cheap ARM hardware,
>but it shows that Apple is pushing hard on exploiting
>multi-core and GPGPU in the apps that need it.

When I've talked to folks at Apple, they seem to have a fairly different perspective. They see a lot of code where performance matters and OpenCL won't be adopted for a while. It takes time. And not all things are suitable for OpenCL.

>>And single threaded performance always matters. There's a >reason why Nvidia's Tesla servers are using Nehalem...
>
>If that were true, Apple wouldn't be anything like as
>profitable as it is. They've been making a ton of money
>the last few years selling fairly slow iMacs and MacBooks
>and Mac Minis. And even slower iPads. There are
>markets where single-thread performance matters a lot
>(and HPC is very definitely one such); and others where
>it really doesn't, as long as you can jump a fairly low bar.
>Lots of people are using web browsers on 800MHz ARM >smartphones quite happily.

I never said it was the most important. Apple probably looks at form factor as a first order constraint, and then performance and efficiency. If you look at the MBA, they use slow CPUs because Intel/AMD didn't have anything that was sufficiently low power (for the form factor) and better in perf or power.

But I think within the context of a given form factor/power envelop, performance does matter. Performance is often interchangeable with efficiency to a certain extent.

I agree that nobody is really saying: "Give us the best performance and we'll design a cooler, PSU, case, etc. for it." But performance really does matter.

>Another point about your article is that probably very few
>of the apps these days have much assembler in them, so
>there would be fewer roadblocks in moving OSX apps
>from x86 to ARM than there were in the transition from
>68K to PowerPC, or even PowerPC to x86.

Yes, that's true. I just assume for the sake of argument that migrating the code is pretty simple. I know Apple is careful about not getting locked in.

David