Comments on: Hell Freezes Over as AMD and Intel Come Together for x86

By: emerth

emerth — Thu, 17 Oct 2024 07:08:05 +0000

@Kevin G, the timelines were as you say. It was competition. If it is to be x86 vs ARM vs bespoke hyperscaler designs, then AMD should just bite the bullet and stage a hostile takeover of Intel.

I don’t see cooperation working beyween the two. Intel’s culture of long sharp knives militates against it.

My $0.02.

By: Kevin G

Kevin G — Wed, 16 Oct 2024 14:21:56 +0000

@emerth While the competition did implement those extension, they did so after they were announced with their competitor having a clear time-to-market advantage. AMD’s implementation of AVX-512 comes five years after Intel ship their first units capable of executing those instructions. The Pentium 4 didn’t gain 64 bit support two years after AMD started shipping their Athlon 64’s. Intel was developing their own 64 bit extensions for x86 but it was Microsoft that put their foot down in that they’d only develop for a single 64 bit extension and they had already committed to AMD’s.

What this collaboration does is hopefully more simultaneous release of new x86 extensions and that they’ll be coherent in their implementation between vendors. This is just as much about appeasing developers as it is being competitive with other architectures. There is a lot of new extensions coming that make this rather important. APX doubles the number of GPR to 32 which brings x86 in-line with many RISC architectures. x86S attempts a full 64 bit implementation. Intel has various AVX10 proposals which I’m personally hoping this collaboration kills several poor choices in it. While not actively being considered, it would be nice to have some consistency for accelerators and heterogeneous compute facilities.

By: emerth

emerth — Wed, 16 Oct 2024 12:49:02 +0000

I dunno about Hell freezing over. They license eachothers tech a lot. Example: Intel EMT64 was actually AMD x86/64 with a blue Intel sticker on the front. Example: AMD implemented AVX2 & 512 but Intel developed them.

By: Anon

Anon — Wed, 16 Oct 2024 04:25:48 +0000

@ Frank: yes, my understanding re. cross licensing is the same as yours. Moreover microsoft has been the de-facto standard enforcer of x86. For example, it was they who compelled intel to adopt amd64. They will generally discourage fragmentation as it makes life harder for windows. As a big buyer of server chips, google has similar leverage. This consortium is simply broadening the tent and making it official. Of course AMD’s increasing market share has precipitated this: x86 standardization was a non-issue in 2015, when AMD was single digit market share. Intel had no incentive to support AMDs ISA extensions, and AMD had no choice but to implement intel’s spec, years after they made it public (see AVX2,AVX512 etc). Today, google and ms would prefer their server fleet to have less isa fragmentation. Hence this. I don’t think ARM has anything to do with this. AMD’s market share gains in server is sufficient enough

By: CS

CS — Wed, 16 Oct 2024 00:48:29 +0000

Seems this is a response to Apple’s M processors actually being competitive. Apple had to drop a ton of legacy support to pull it off though (which they then later supported rather well through an emulation layer). Intel/AMD don’t have that luxury … yet. I imagine Intel/AMDs collab here will most likely be the opposite of technical progress, but in an underlying agreement on which legacy parts of the ISA to BOTH stop shipping. Just looking at the other members of the advisory list seems to support this, because OS vendors (Microsoft/Red Hat/Google) will need to pull out those code paths or re-architect under the assumption that they’re gone.

Apple has the advantage (vertical integration) of designing its OS and Processor to work intimately together and only worry about that. Intel/AMD/Microsoft/Google/etc do not.

For reference, not an Apple apologist here. I actually think they’re quite awful.

By: TheTamago

TheTamago — Tue, 15 Oct 2024 23:53:13 +0000

I am getting a big laugh out of this joint effort after the 1990s early 2000s for the hell between them. It still hurts in the OEM space other than the Enterprise arena for AMD. Still, the rise of ARM’s easier performance characteristics and when RISC was still the mighty Enterprise architecture from the ancients. It all falls down on the OS, HV, Apps to dictate what can do better. Still Perf/W might be the standard of choice?!