By: Linus Torvalds (torvalds.delete@this.linux-foundation.org),
Room: Moderated Discussions
Michael S (already5chosen.delete@this.yahoo.com) on February 21, 2019 8:35 am wrote:
>
> A counter-argument is that internally N1==A76 so Snapdragon 855 is not
> far away from what you're asking for.
I do agree that the ARM laptops may make this all work out on the development side. Whether from the PC side ("WARM") or Apple cutting their laptops over.
It doesn't have to be the same core, just "close enough" and with good enough performance and resources to work as a development box.
Because it's just very frustrating to chase bugs that you can't reproduce locally. Particularly if they are in random infrastructure that you don't know (and perhaps not even "bugs" - just different behavior in basic libraries, perhaps just due to implementation-defined differences).
When everything works, switching from one architecture to another should be a walk in the park. It's just that people seem to unrealistically assume that things work, when in reality, most development effort - even if you have just a single platform - is exactly in that "oops, it didn't work as intended" situations.
Which is why I've been so adamant about asking for ARM developer boxes (not "development boards" - those are for hardware people).
That said, I still strongly suspect that any "hyperscaling" 64-128 core system will be very underwhelming. It may be just good enough on Specrate style "nothing shared" benchmarks, and yes, you have the traditional network processor loads etc, but I think people always underestimate how big of a jump it really is from "look, we know how to make 4 cores work" to "look, we can cram a lot of cores on a die" and then to actually having something that truly scales.
You simply need many many generations. And even then it's hard. So says a lot of mostly dead companies. Right now, ARM doesn't have even a single generation of server parts out, and they are pushing the hyperscaling story? Does that really make sense to anybody?
(Yes, yes, I realize it makes tons of sense to all those people who already believe in the "sea or cores" fairytale despite not having any evidence of that ever working. But there's no arguing with delusion)
Linus
>
> A counter-argument is that internally N1==A76 so Snapdragon 855 is not
> far away from what you're asking for.
I do agree that the ARM laptops may make this all work out on the development side. Whether from the PC side ("WARM") or Apple cutting their laptops over.
It doesn't have to be the same core, just "close enough" and with good enough performance and resources to work as a development box.
Because it's just very frustrating to chase bugs that you can't reproduce locally. Particularly if they are in random infrastructure that you don't know (and perhaps not even "bugs" - just different behavior in basic libraries, perhaps just due to implementation-defined differences).
When everything works, switching from one architecture to another should be a walk in the park. It's just that people seem to unrealistically assume that things work, when in reality, most development effort - even if you have just a single platform - is exactly in that "oops, it didn't work as intended" situations.
Which is why I've been so adamant about asking for ARM developer boxes (not "development boards" - those are for hardware people).
That said, I still strongly suspect that any "hyperscaling" 64-128 core system will be very underwhelming. It may be just good enough on Specrate style "nothing shared" benchmarks, and yes, you have the traditional network processor loads etc, but I think people always underestimate how big of a jump it really is from "look, we know how to make 4 cores work" to "look, we can cram a lot of cores on a die" and then to actually having something that truly scales.
You simply need many many generations. And even then it's hard. So says a lot of mostly dead companies. Right now, ARM doesn't have even a single generation of server parts out, and they are pushing the hyperscaling story? Does that really make sense to anybody?
(Yes, yes, I realize it makes tons of sense to all those people who already believe in the "sea or cores" fairytale despite not having any evidence of that ever working. But there's no arguing with delusion)
Linus
