AMD, Intel Mobilize Major Tech Companies To Extend X86 Architecture (Interview)
Posted by TwelveSilverSwords@reddit | hardware | View on Reddit | 47 comments
3G6A5W338E@reddit
They are afraid. They are very afraid.
Of RISC-V.
SirLimonada@reddit
definitely this is the year of risc-v
3G6A5W338E@reddit
2024 was cool, but I'd point at 2025.
That's the year where we'll start seeing the larger microarchitectures show up in actual chips. Ascalon, P670, Ventana V2 and such.
SignalButterscotch73@reddit
I'm probably too ignorant of cpu design to understand this fully but I'm guessing this means that AMD won't be a gen behind on things like AVX and SSE instructions like in the past and that mostly ignored instructions like 3DNow! won't happen again, or is this completely different from ISA cooperation?
Azzcrakbandit@reddit
Isn't amd already ahead with avx? Intel had to remove it due to efficiency cores not supporting it.
advester@reddit
Intel has designed the future standard for AVX in AVX10. It can't become the new baseline unless AMD implements it too.
Tuna-Fish2@reddit
AVX10.1/256 is the 256-bit versions of all the AVX-512 instructions. AMD already implements them.
nanonan@reddit
Sure, and Intel wants to implement them in a slightly different way, which is a problem that needs resolution.
Tuna-Fish2@reddit
No, they don't.
The intel proposal for AVX10.1/256 is literally just all the instructions that they already implement, except leave out the 512-bit variants. Every CPU that can execute AVX512VL (+ all the sapphire rapids extensions) is already capable of running all AVX10.1/256 instructions.
SignalButterscotch73@reddit
AVX512 is such a shitshow I don't know who's ahead or behind. Intel have dumped it on the consumer side but both of them have it in server.
Thers around a dozen or more different subsets within the AVX512 instruction set and I have no clue who has what on what cpu. I vaguely remember a comment saying that some of the subsets are even incompatible with each other, no clue if that's true but I'm temped to believe it.
Thankfully nothing I use needs it, so its not something I need to worry about yet.
Azzcrakbandit@reddit
Currently, amd is ahead and has been since the ryzen 7000 series. The zen 4 and 4c onwards are similar enough that both cores support avx 512 while intels p and e cores aren't, hence why the e cores lack avx 512. When intel did have it, you either had to disable avx 512 to use all the cores or disable all the e cores to use avx 512.
kyralfie@reddit
This thread is ridiculous. You are talking about consumer parts and get a 'counter-argument' about server parts. And you are the one downvoted, smh.
SignalButterscotch73@reddit
The current gen Xeon chips (Sapphire Rapids) all have AVX512 support its just the consumer parts that don't.
Azzcrakbandit@reddit
Ah, so something irrelevant to 99% of people.
SignalButterscotch73@reddit
Dude, AVX512 is irrelevant to 99.99% of people 🙄
Azzcrakbandit@reddit
Except when it isn't. Amd wouldn't put something that consumes significant die space in consumer cpus if 99.99% of consumers don't use or need it.
SignalButterscotch73@reddit
Chiplet's. Everything AMD within a single generation uses the exact same chiplet's. AMD didn't decide to add AVX512 for the benefit of consumers. Every single AMD cpu is designed for the data centre, Ryzen are the rejected chips left over from Epyc binning.
monocasa@reddit
The APUs will have it and they're not chiplets.
Azzcrakbandit@reddit
Every ryzen before 7000 didn't have it even though they made data centre chips with ryzen before.
Dark_Souls_VII@reddit
It gets even more funny. I transcode video with SVT-AV1 and that can be AVX512 accelerated. Unfortunately my 9700X (ZEN5) is so memory bandwidth limited it barely is an improvement over just using AVX2 in that task for me.
Arbiter02@reddit
Interesting… I presume Epyc/threadripper would be a better fit for you then? Of course one of those chips alone could set you back about as much as an entire AM5 PC alone lolÂ
Dark_Souls_VII@reddit
I have EPYCs at work. Because of electricity cost alone I would not run something big at home. Ryzen is neat because it uses the same chiplets. So it's good for prototyping and playing a game from time to time.
the_dude_that_faps@reddit
To be fair, a lot of the benefits of avx512 are aside from the vector width. Vector width is irrelevant for RPCS3, for example.
klapetocore@reddit
most programmers use AVX512F and AVX512BW, so you usually check for these 2. the actual problem with AVX512 was that intel's implementation was throttling the clocks to use it and had a lot of cycles of lag to boot it up. Unless you knew how to use it correctly, it did not make sense to use it because you would end up with worse performance because of the downclock, so most of the programmers did not use it. Amd on the other hand countered a lot of these with their implementation on zen5 and it makes more sense to use it on these processors.
Exist50@reddit
The frequency thing was basically fixed with Ice Lake.
BlackenedGem@reddit
It was basically "let's have a mediocre implementation for the first gen so it's supported for longer in hardware. Then when we get a logic shrink at 10nm all of these frequency problems go away and we get the first proper implementation".
Except they then got stuck at 14nm for 5 years.
Exist50@reddit
Eh, it was also a design solution. With Sunny Cove, they basically added a heuristic for power intensive vs not op detection. With Golden Cove, they have a whole-ass microcontroller for this and other things.
Jeep-Eep@reddit
I suspect it's also a tool for both to breath down folks like MS's necks about problems like schedulers.
Hen-stepper@reddit
It sounds like a good news. I don't know why I'm cheering for X86. It seems like ARM directly threatens PC building and gaming as a whole. Having strong, separate components is fun and improves gaming performance over say Apple laptops.
Reversi8@reddit
There's nothing about ARM that would prevent separate components though.
Hen-stepper@reddit
I guess there's nothing forcing ARM to use unified memory or soldered parts like Apple does, but that's the popular implementation.
When Apple used IBM RISC processors all of these parts were separated. It doesn't seem like we are going back to that.
imaginary_num6er@reddit
When is Nvidia going to lead the alliance of non-x86 architectures?
BatteryPoweredFriend@reddit
Whenever someone uploads their fanfic of it onto AO3.
rebelrosemerve@reddit
I can write something but I already made them disappeared.
AbheekG@reddit
They tried to by buying ARM but everyone else panicked so hard they realized it’s not happening.
xNOOPSx@reddit
Given their love of proprietary all things, I think the response from everyone was warranted.
3G6A5W338E@reddit
It gave a nice boost to RISC-V, which cannot be "bought out" by any single company.
This is particularly true now that RISC-V has become huge while the foundation has remained diverse in its composition.
monocasa@reddit
That was a fair thing to panic over, IMO.
MixtureBackground612@reddit
It's gonna take 8 years for result?
Arbiter02@reddit
Stuck in the past as usual, fortunately for them Microsoft is too
3G6A5W338E@reddit
Unfortunately for them, Microsoft has been putting the work.
Not only they brought Windows for ARM to an usable state, and successfully implemented x86 emulation thus enabling a transition away from x86 which would otherwise be impossible.
But they are also have been quite active, for many years now, on RISC-V specification work groups.
Arbiter02@reddit
Microsoft drops anything that isn't making them fat stacks of cash almost immediately, not a hard trend to find in their history. They like to keep their fingers in lots of pies, but if it doesn't see wide adoption or costs them money without a return you can bet it'll get dropped almost immediately. "Usable state" is also a MASSIVE stretch. "Barely Functional" would be more accurate from the long-term reviews I've seen, and it's still outright laughable compared to MacOS and mobile platforms. And to be perfectly fair, apple and mobile devs have the power to say "Get with the program or get fucked" to software developers, while Microsoft simply has no such power. That doesn't change the fact that ARM on windows is playing 8th fiddle to a thousand other projects though.
I've yet to see RISC-V manifest in anything but some insanely niche scenarios. Until something drastically changes it's simply never going to overcome ARM, it just isn't in any meaningful end-user hardware at all. Very interesting instruction set but it isn't taking off anytime soon at the level that windows operates in especially. Microsoft participating in discussions around it isn't indicative of meaningful support, it's just expected of them as a large tech company, particularly a cloud provider like themselves that's interested in RISC-V for it's applications in the supercomputer/server environment
braaaaaaainworms@reddit
I have an ARM laptop and windows on it is far more usable than you think
LoveBigCOCK-s@reddit
The Future is now, old man.
Intel last gen generally x86 with removed hyperthead and low TDP
performance did not come close to ARM64.
Now Intel marketing spreading a lot of news, and articles to call for confidence from stock inverstor
SherbertExisting3509@reddit
I hope this is the start of improvements being made to the x86-64 ISA. adoption of x86s would be good start and I'm excited to see what comes of this partnership. AVX10 adoption by both companies would be a good option.
Intel and AMD should collaborate on creating an X86 equivalent of SVE (variable vector length SIMD instruction)
3G6A5W338E@reddit
nekogami87@reddit
It's just an advisory board, important to get the vision of the future of x86, but both will probably still be in charge of how to do something and when. Imo, it's more going to be about, how to get rid of old, costly instructions but still ensure backward compatibility. OR even decide "you might wanna use Linux for very older stuff now".