By: Ungo (a.delete@this.b.c.d.e), May 23, 2022 12:27 am
Room: Moderated Discussions
Jan Wassenberg (jan.wassenberg.delete@this.gmail.com) on May 22, 2022 11:11 pm wrote:
> But given that we are moving into a world with more
> heterogeneity (whether we like it or not), isn't that an occasion to perhaps adapt what we've been doing?
Who says we're moving into a world with more ISA-level heterogeneity in general purpose personal computers? Performance and efficiency heterogeneity is clearly here to stay, but in application processor cores, ISA heterogeneity doesn't seem to have a future.
Consider that Alder Lake's ISA heterogeneity proved to be such a problem that Intel decided to patch out a huge chunk of it by disabling AVX512 in AL's P cores. Doesn't that tell you something?
So I don't see AL as a trendsetter. I think it's something unique to Intel's current circumstances. A few years ago, Intel must have realized they needed a high performance heterogenous chip yesterday, and the fastest path to get one to market was to repurpose an Atom core as an E core. If Intel is being run well, in the near future they should begin shipping heterogenous chips whose cores are far better matched to each other than Gracemont and Golden Cove.
It's not just x86 either, there have been a few painful object lessons in the Arm space too. Today's standard practice is to pair Arm cores with identical (or nearly so) ISA features.
> But given that we are moving into a world with more
> heterogeneity (whether we like it or not), isn't that an occasion to perhaps adapt what we've been doing?
Who says we're moving into a world with more ISA-level heterogeneity in general purpose personal computers? Performance and efficiency heterogeneity is clearly here to stay, but in application processor cores, ISA heterogeneity doesn't seem to have a future.
Consider that Alder Lake's ISA heterogeneity proved to be such a problem that Intel decided to patch out a huge chunk of it by disabling AVX512 in AL's P cores. Doesn't that tell you something?
So I don't see AL as a trendsetter. I think it's something unique to Intel's current circumstances. A few years ago, Intel must have realized they needed a high performance heterogenous chip yesterday, and the fastest path to get one to market was to repurpose an Atom core as an E core. If Intel is being run well, in the near future they should begin shipping heterogenous chips whose cores are far better matched to each other than Gracemont and Golden Cove.
It's not just x86 either, there have been a few painful object lessons in the Arm space too. Today's standard practice is to pair Arm cores with identical (or nearly so) ISA features.