Article: AMD's Mobile Strategy
By: Exophase (exophase.delete@this.gmail.com), December 15, 2011 3:15 pm
Room: Moderated Discussions
Linus Torvalds (torvalds@linux-foundation.org) on 12/15/11 wrote:
---------------------------
>Why do you even talk about ARM64? We have no idea how
>well it will work, and how good/bad it will be. It looks
>like a largely traditional RISC architecture, and we know
>just how well those worked.
>
>For all we know, it will have the same problems with
>huge I$ footprints etc. There's nothing special about it,
>and it throws out everything ARM learnt from Thumb2.
>
>Sure, it may be easy to decode, but so was alpha. Nobody
>cares. That's not where the problems spots tend to be.
>
>Linus
I'm talking about ARM64 because it'll be four-way first, and because we're comparing instruction sets, not uarch efficiency. Your response makes it sound as if we know nothing about ARM64 when we know pretty much everything about it. Have you read the spec? If you haven't maybe you should. It doesn't seem like much of a "traditional RISC architecture" to me; it reduces some of the specialties of traditional ARM but does so without fully removing them while adding several new enhancements.
And really, how well traditional RISCs have done in general is immaterial. The discussion at hand is amount of work done per instruction and nothing else. Stop diverting.
Thumb-2 exists to decrease code footprint on embedded profile (Cortex-M and Cortex-R) series processors, that have to run code from limited flash. What ARM learned from Thumb-2 on Cortex-A series processors is it doesn't improve performance at all, despite prevailing myths about code density and performance in typical programs on modern platforms. Why they included it in ARM-v7a is a mystery. Why they abandoned it or anything like it in AArch64 is not.
---------------------------
>Why do you even talk about ARM64? We have no idea how
>well it will work, and how good/bad it will be. It looks
>like a largely traditional RISC architecture, and we know
>just how well those worked.
>
>For all we know, it will have the same problems with
>huge I$ footprints etc. There's nothing special about it,
>and it throws out everything ARM learnt from Thumb2.
>
>Sure, it may be easy to decode, but so was alpha. Nobody
>cares. That's not where the problems spots tend to be.
>
>Linus
I'm talking about ARM64 because it'll be four-way first, and because we're comparing instruction sets, not uarch efficiency. Your response makes it sound as if we know nothing about ARM64 when we know pretty much everything about it. Have you read the spec? If you haven't maybe you should. It doesn't seem like much of a "traditional RISC architecture" to me; it reduces some of the specialties of traditional ARM but does so without fully removing them while adding several new enhancements.
And really, how well traditional RISCs have done in general is immaterial. The discussion at hand is amount of work done per instruction and nothing else. Stop diverting.
Thumb-2 exists to decrease code footprint on embedded profile (Cortex-M and Cortex-R) series processors, that have to run code from limited flash. What ARM learned from Thumb-2 on Cortex-A series processors is it doesn't improve performance at all, despite prevailing myths about code density and performance in typical programs on modern platforms. Why they included it in ARM-v7a is a mystery. Why they abandoned it or anything like it in AArch64 is not.