Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads

[Tigerick]

As Hot Chips 34 starting this week, Intel will unveil technical information of upcoming Meteor Lake (MTL) and Arrow Lake (ARL), new generation platform after Raptor Lake. Both MTL and ARL represent new direction which Intel will move to multiple chiplets and combine as one SoC platform.

MTL also represents new compute tile that based on Intel 4 process which is based on EUV lithography, a first from Intel. Intel expects to ship MTL mobile SoC in 2023.

ARL will come after MTL so Intel should be shipping it in 2024, that is what Intel roadmap is telling us. ARL compute tile will be manufactured by Intel 20A process, a first from Intel to use GAA transistors called RibbonFET.





Intel Core Ultra 100 - Meteor Lake



As mentioned by Tomshardware, TSMC will manufacture the I/O, SoC, and GPU tiles. That means Intel will manufacture only the CPU and Foveros tiles. (Notably, Intel calls the I/O tile an 'I/O Expander,' hence the IOE moniker.)

 

[OneEng2]

Phoronix has a piece on this: https://www.phoronix.com/review/amd-epyc-9755-avx512

256double pump works well enough for many applications.

Looking at the details, where AVX512 made a big difference, the delta for full 512 vs double pumped 256 was decent (say ~25% on average). Unfortunately, lots of benchmarks were included where AVX512 didn't make much difference at all (<10%). This made the geo mean only 12% more on all the tests done between 512 and 256.

Still, your point is valid. 256 double pumped does give the majority of the value. I do wonder how much more die space it took to have full 512 though?

The problem is AMD doens't have the resources to put into building the Software ecosystem for x86 it is still relying on Intel to do the gruntwork it is the reason they are failing in GPU race cause they can't magically run CUDA(zluda exists) like they can with x86 and AVX10 is a good step from point of view of many SW people the only issue is the data path

I think AVX10 extensions are just more of Intel trying to gain some wins in a few early benchmarks on release day. The inclusion of AVX512 backwards compatibility (through double pumping 256bit registers) will likely be the most impactful. I think the extensions will be largely unused for quite some time. We will see.

They threw the towel cause their P/E core dilema

I don't think it was a dilemma at all. I think it was economics. Adding AVX512 added die space. Lion Cove was already over-budget size wise (my guess), and the added front end and retire work needed in addition to the 512b data path in Skymont was simply out of the question. This is also why SMT was left off of Skymont. I think Intel is paying lots of money to TSMC for those N3B tiles, and I don't think there was any appetite at all for an even larger die.

I understand the sentiment that Intel may have slowed down innovation for consumers, but from a business perspective it makes perfect sense as Zen4/5 pays the early adopter fees of including AVX-512 in their client architectures while Intel only spends the extra effort for AVX-512 in consumer products when it actually matters.

Intel paid the fees already. AMD just did it better and Intel stopped due to economics on desktop (the juice wasn't worth the squeeze in desktop).

I have it on good authority that we can expect both Darkmont and Panther CoveX to have AVX10 support (and backwards compatibility with AVX512). Panther Cove X will have SMT, but Darkmont will not.

Lol. Years of Intel failure and AMD managed to capture only around ~20% client market share. They're total dumb@$$. AMD is subpar. And with the launch of ARL-H, they're pretty much out of the picture (except for some tiny verticals).

LOL indeed. AMD took home more server revenue than AMD last quarter sucking up the most profitable market and fastest growing market in x86. THIS was the strategy of the Zen architecture all along. Additionally, AMD has been constantly gaining market share from Intel across all x86 markets.

ARL-H occupies a very small part of the x86 market .... and it won't be without competition. Stirx Halo will be released in early 2025 I believe. And AMD will be competing with Intel's 3nm chips with 4nm chips (which are less expensive).

How do you figure AMD is "Out of the picture"? Seems more than just wildly optimistic to think so IMO.

Nope. They aren't backward/forward compatible. AVX-512 code can't just be recompiled to work with AVX-10 and vice versa. They need to be recoded. Time to say goodbye to AMD's client implementation of AVX-512.

Only the AVX-10 extensions. You should read the information on AVX10.

Anyways, it's a ton of work for a edge-case feature on a consumer only product. If they want to add AVX-512 they should add it to Darkmont for server, but it looks like they aren't.

I hear they are .... well, AVX10 and backwards compatible double pumped AVX512 anyway . Even if they didn't, from what I can tell for the Darkmont use case in Clearwater Forest, HPC and other high CPU use isn't what it is meant for. That will be Diamond Rapids.... which I am guessing wont have any Darkmont cores at all, but rather Panther CoveX cores.

 

[ajsdkflsdjfio]

...wat?

AVX512 may be niche but let's not rewrite history.

Intel Core i9-14900K, i7-14700K and i5-14600K Review: Ryzen X3D Stays On Top

Mid-range gaming has a new champion.

www.tomshardware.com

You're right, maybe more like 60-70% in productivity and 90% in gaming. I remembered gaming performance and confused it with productivity, still overall beats Zen 4 handily but with the caveat of 300W power consumption.

Not comparing with X3D since X3D is separate from the discussion of the core architecture.

I hear they are .... well, AVX10 and backwards compatible double pumped AVX512 anyway . Even if they didn't, from what I can tell for the Darkmont use case in Clearwater Forest, HPC and other high CPU use isn't what it is meant for. That will be Diamond Rapids.... which I am guessing wont have any Darkmont cores at all, but rather Panther CoveX cores.

https://twitter.com/x/status/1870437498037415956

DK how accurate this is but seems like no AVX10 on darkmont/clearwater forest. Would be amazing if it did have AVX10 just to be the cherry on top for clearwater forest.

 

[DrMrLordX]

Intel Core i9-14900K, i7-14700K and i5-14600K Review: Ryzen X3D Stays On Top

Mid-range gaming has a new champion.

www.tomshardware.com

You're right, maybe more like 60-70% in productivity

The link you just provided showed the 7950X winning MT productivity geomean of tests.

The Ryzen 9 7950X is ~2% faster than the Core i9-14900K in our cumulative measurement of multi-threaded workloads.

 

[ajsdkflsdjfio]

The link you just provided showed the 7950X winning MT productivity geomean of tests.

TLDR:
Tomshardware: 61%
Anandtech: 64%
Techpowerup: 79%

You know if you actually stopped and looked at the data the raptor lake is winning 60-70% of the time, it's just that in certain benchmarks they lose, they lose by a wider margin due to a lack of certain features so the geomean is lower.
THW rendering: 6/11 (not including single-core)
THW encoding: 4/5 (no single-core tests)
THW miscellaneous: 3/5 (not including single-core)
Overall: 13/21 = 61% 13900k wins
If you include single core tests the 13900k wins in 69% of tests.

DK why you ignored single thread performance either since it's still arguably relevant for the discussion of productivity performance and especially for the original discussion of core architecture vs core architecture which spawned these further comparisons.


Other sites show a similar trend of 60-70% wins in productivity. Anandtech for example(whose forums you all have posted thousands of comments on):

Intel Core i9-14900K, Core i7-14700K and Core i5-14600K Review: Raptor Lake Refreshed

www.anandtech.com

productivity page 1: 7/12 58%
prod page 2: 15.5/18 85% (one tied)
prod page 3: 6/11 55%
prod page 4: 13/17 76%
prod page 5: 4/13 30%

Overall productivity tests: 45.5/71 = 64%, even including the 13 AI inferencing tests which benefit heavily from AVX-512.

Intel Core i9-14900K Review - Reaching for the Performance Crown

Core i9-14900K is Intel's new flagship with clock speeds of up to 6 GHz. It's actually clocked even higher than the 13900KS, thanks to an extra 100 MHz when more than two cores are active. Our review confirms that Raptor Lake Refresh is amazing for both applications and gaming, if you can live...

www.techpowerup.com

Synthetic: 2/3
Rendering: 3/4
soft dev: 2/2
web use: 3/3
AI: 1/3
science/simulation: 3/3
office and misc productivity: 5/6
more misc: 1/3
server and virtualization: 5/5
compression/encryption: 1.5/4 (won in one version of benchmark, lost in other)
(skipped emulation tests, but it's 1/2 if you care)
media encoding: 4/4
Total productivity: 31.5/40 = 79%

More tests (too lazy to count but all show same results or better):

Intel Core 14th Gen Processors Content Creation Review

The new Intel Core 14th Gen processors are a refresh update of the previous 13th Gen processors. But, how much faster are they for content creation workflows?

www.pugetsystems.com

https://www.pcmag.com/reviews/intel-core-i9-14900k

I didn't cherry pick these sites either, they are just some of the top results that come up when searching 13th/14th gen productivity and who have a decent suite of productivity tests.

Not even saying raptor lake is good, if I were to build a pc anytime after zen3x3d release or especially zen4 release I would have chosen zen4. Raptor lake is shit because it's literally a copy paste of alder lake but with more power instead of any actual innovation. It turns out this is enough to beat Zen4 but it still isn't a good product for the average consumer since that extra 5-10% of performance comes with extra heat and possibly system instability.

No need to lie about performance though or about core architecture when were are having a purely theoretical discussion. Both things can be true at once, raptor lake can be bad vs zen4 while still having better perf.

 

[Thunder 57]

You know if you actually stopped and looked at the data the raptor lake is winning 60-70% of the time, it's just that in certain benchmarks they lose, they lose by a wider margin due to a lack of certain features so the geomean is lower.
THW rendering: 6/11 (not including single-core)
THW encoding: 4/5 (no single-core tests)
THW miscellaneous: 3/5 (not including single-core)
Overall: 13/21 = 61% 13900k wins
If you included single core tests the 13900k wins in 69% of tests.

DK why you ignored single thread performance either since it's still arguably relevant for the discussion of productivity performance and especially for the original discussion of core architecture vs core architecture which spawned these further comparisons.

Other sites show a similar trend of 60-70% wins in productivity. Anandtech for example(whose forums you all have posted thousands of comments on):

Intel Core i9-14900K, Core i7-14700K and Core i5-14600K Review: Raptor Lake Refreshed

www.anandtech.com

7/12 58%

Intel Core i9-14900K, Core i7-14700K and Core i5-14600K Review: Raptor Lake Refreshed

www.anandtech.com

15.5/18 85% (one tied)

Intel Core i9-14900K, Core i7-14700K and Core i5-14600K Review: Raptor Lake Refreshed

www.anandtech.com

6/11 55%

Intel Core i9-14900K, Core i7-14700K and Core i5-14600K Review: Raptor Lake Refreshed

www.anandtech.com

13/17 76%

Intel Core i9-14900K, Core i7-14700K and Core i5-14600K Review: Raptor Lake Refreshed

www.anandtech.com

4/13 30%

Overall productivity tests: 45.5/71 = 64%, even including the 13 AI inferencing tests which benefit heavily from AVX-512.

Lol you are comparing win percentage as if it were a performance advantage. And then you are twisting what does or doesn't count to fit your narrative. Then you go on to say it was the best despite drawing ~300W when the review you linked mentions 428W. I'm not sure how to comment on that.

 

[ajsdkflsdjfio]

Lol you are comparing win percentage as if it were a performance advantage. And then you are twisting what does or doesn't count to fit your narrative. Then you go on to say it was the best despite drawing ~300W when the review you linked mentions 428W. I'm not sure how to comment on that.

When did I say it was the best? When did I say win percentage was performance advantage? When did I twist what does or doesn't count? I included all the possible tests on all those sites? I even further addressed that Raptor lake was actually bad IMO because of the power draw?

I'll give you that last one since you posted your response before I edited my post to further elaborate on my thoughts about raptor lake as a product. Even so, you shouldn't assume I am claiming Raptor lake is THE BEST, just because I give solid evidence indicating it was more performant... You need to take a lesson in nuance, not everything is black and white. Just because I am saying raptor lake was more performant than Zen4 doesn't mean I am saying that it was a better product.

My original post literally said 60-70% in productivity and 90% in gaming. Why would I say 90% in gaming if I were talking about performance advantage? Do you think I am claiming raptor lake has a 1.9x gaming performance advantage? Or do you think I am claiming raptor lake has 90% or .9x of the gaming performance of Zen4? If you actually used your brain for a second instead of assuming I was riding the raptor lake train, then you would realize that I was talking about win percentage and only that. My only point was that raptor lake performed better in MOST productivity apps, and MOST games.

And to follow up on the argument that Zen4 is more performant on average (not win percentage). In every source I listed besides Tomshardware, the geomean of tests would indicate the 14900k as the winner(maybe not anandtech because of all the aI tests, i'm not sure though since the rest of the tests have a solid advantage for raptor). Regardless this is a moot point.

 

[hemedans]

https://twitter.com/x/status/1873513849200230803

Lunar Lake 20/15/10W perfomance

 

[Thunder 57]

When did I say win percentage was performance advantage?

You didn't. It just seemed worded odd to me. You said 60-70% win percentage. That's just not something I've ever heard before. It is a stupid metric because you go to another review and they run some more or different becnhmarks and the number can completely flip. However I just checked the thread and apparantly this isn't a RPL thread so I'll drop it.

Nice edit though, except for saying you didn't cherry pick sites and just used the top search results. I think we all know how lowerbenchmark is typically a top result.

 

[DrMrLordX]

You know if you actually stopped and looked at the data the raptor lake is winning 60-70% of the time, it's just that in certain benchmarks they lose, they lose by a wider margin due to a lack of certain features so the geomean is lower.

Okay, but if the losses are of such a small margin on other applications that an old previous-gen CPU can catch up enough to win the geomean by 2%, that probably means the 7950X is competitive. If you happen to need the CPU for one of those benches where the margin is not in Intel's favor . . .

DK why you ignored single thread performance either since it's still arguably relevant for the discussion of productivity performance and especially for the original discussion of core architecture vs core architecture which spawned these further comparisons.

In actual applications where we can see both ST and MT data, ST data is no longer relevant in the vast majority of cases where productivity is concerned. Very few performance-sensitive applications spawn a single thread anymore.

To return to the post to which I initially replied, it's not a good idea to make blanket statements like "Alder Lake/Raptor Lake are better applications processors than Zen4" (paraphrasing). It's clearly preposterous in the case of Alder Lake, and Raptor Lake was only able to achieve narrow victory despite adding quite a few more cores and raising power limits. Never mind that Raptor Lake was also critically flawed. It's best not to use such a . . . questionable observation as the underpinning for a broader argument.

 

[Meteor Late]

Okay, but if the losses are of such a small margin on other applications that an old previous-gen CPU can catch up enough to win the geomean by 2%, that probably means the 7950X is competitive. If you happen to need the CPU for one of those benches where the margin is not in Intel's favor . . .



In actual applications where we can see both ST and MT data, ST data is no longer relevant in the vast majority of cases where productivity is concerned. Very few performance-sensitive applications spawn a single thread anymore.

To return to the post to which I initially replied, it's not a good idea to make blanket statements like "Alder Lake/Raptor Lake are better applications processors than Zen4" (paraphrasing). It's clearly preposterous in the case of Alder Lake, and Raptor Lake was only able to achieve narrow victory despite adding quite a few more cores and raising power limits. Never mind that Raptor Lake was also critically flawed. It's best not to use such a . . . questionable observation as the underpinning for a broader argument.

ST is relevant, not by itself, but because people use either ST or embarrassingly parallel MT benchmark workloads that don't represent most real world workloads.
What I mean by this is that ST may not be relevant, but two core or 4 core performance absolutely is. That's why ST benchmark scores are absolutely relevant, when given ST or MT options to choose from.

 

[Asterox]

Arrow Lake Core Ultra 9 285K, so that there is no confusion.

https://twitter.com/x/status/1873698469724209485

 

[Joe NYC]

Arrow Lake Core Ultra 9 285K, so that there is no confusion.

https://twitter.com/x/status/1873698469724209485

Just going through a rough patch.

 

[Schmide]

Nope. They aren't backward/forward compatible. AVX-512 code can't just be recompiled to work with AVX-10 and vice versa. They need to be recoded. Time to say goodbye to AMD's client implementation of AVX-512.


And with a very performant 285H/265H lineup, they're pretty much on solid ground.

Avx 10.x is intel's way of repackaging the functionality of avx512 to 256b/128b processors. Avx512 provided a lot of the missing functionality avx2 needed. (BF16, VNNI, IFMA, etc) A lot of these operators have 256b/128b pathways within the 512b standard, avx10.x just delineates the lower bit functions such that ecores can use them without providing the 512b path.

For the most part they are backwards compatible as the avx10.x can be a subset of the latest avx512 standard.

 

[OneEng2]

Intel Core i9-14900K, i7-14700K and i5-14600K Review: Ryzen X3D Stays On Top

Mid-range gaming has a new champion.

www.tomshardware.com

You're right, maybe more like 60-70% in productivity and 90% in gaming. I remembered gaming performance and confused it with productivity, still overall beats Zen 4 handily but with the caveat of 300W power consumption.

Not comparing with X3D since X3D is separate from the discussion of the core architecture.

https://twitter.com/x/status/1870437498037415956

DK how accurate this is but seems like no AVX10 on darkmont/clearwater forest.

I might have misinterpreted what I was told.

Perhaps it is just Panther CoveX that has AVX10 and SMT.

 

[511]

Avx 10.x is intel's way of repackaging the functionality of avx512 to 256b/128b processors. Avx512 provided a lot of the missing functionality avx2 needed. (BF16, VNNI, IFMA, etc) A lot of these operators have 256b/128b pathways within the 512b standard, avx10.x just delineates the lower bit functions such that ecores can use them without providing the 512b path.

For the most part they are backwards compatible as the avx10.x can be a subset of the latest avx512 standard.

I am glad they killed 128b vector path so we get minimum 256b at least

 

[SiliconFly]

.

 

[ajsdkflsdjfio]

You didn't. It just seemed worded odd to me. You said 60-70% win percentage. That's just not something I've ever heard before. It is a stupid metric because you go to another review and they run some more or different becnhmarks and the number can completely flip. However I just checked the thread and apparantly this isn't a RPL thread so I'll drop it.

Nice edit though, except for saying you didn't cherry pick sites and just used the top search results. I think we all know how lowerbenchmark is typically a top result.

Whether it's a stupid metric or not sure that's debatable, but the results all skew towards one side no matter how many sites you go to, you say they flip, but if you read for more than one second you can see intel usually has the advantage.

Either way I don't give a shit about win percentage either, I just said that 60-70% number to show that raptor lake wins in most productivity apps which it does. DrMrLordX tried to call me out on it by bringing up geomean, without actually looking at my statement or the data.

Yes I didn't cherry pick tests, the only three reviews I chose to count and determine percentages, two were on the first page, and the third (Tomshardware) was the link I originally posted which DrMrLordX responded to. Quit being ignorant.

Okay, but if the losses are of such a small margin on other applications that an old previous-gen CPU can catch up enough to win the geomean by 2%, that probably means the 7950X is competitive. If you happen to need the CPU for one of those benches where the margin is not in Intel's favor . . .

Which was never my intention? I never intended to say 7950x is uncompetitive in every productivity scenario, but that it loses in 60-70% of productivity applications.

In actual applications where we can see both ST and MT data, ST data is no longer relevant in the vast majority of cases where productivity is concerned. Very few performance-sensitive applications spawn a single thread anymore.

Very few performance-sensitive applications spawn a single thread, but there are many applications which only a couple threads/cores and don't fully saturate the CPU in which single core is still relevant, especially if the program doesn't bleed into intel's weaker e-cores in the raptor lake generation.

To return to the post to which I initially replied, it's not a good idea to make blanket statements like "Alder Lake/Raptor Lake are better applications processors than Zen4" (paraphrasing). It's clearly preposterous in the case of Alder Lake, and Raptor Lake was only able to achieve narrow victory despite adding quite a few more cores and raising power limits. Never mind that Raptor Lake was also critically flawed. It's best not to use such a . . . questionable observation as the underpinning for a broader argument.

Raptor lake achieve narrow victory in productivity sure = Raptor lake wins in 60-70% of productivity applications they both mean the same thing. And my argument was never that raptor lake was a good product. I made a general statement for a very small point, everyone just seems to think that I suddenly love raptor lake and think it's a great product since I stated facts about it's performance advantages.

statements like "Alder Lake/Raptor Lake are better applications processors than Zen4" (paraphrasing).

Saying that raptor lake wins in 60-70% of productivity applications doesn't mean that Raptor lake is a better processor, neither did I ever intend to say something like that. My only point was that Raptor lake was competitive in performance to other processors which had AVX-512 enabled, this is my original quote:

Also It's funny how in those graphs rocket-lake and zen-4 demolish Raptor lake without AVX-512 in those workloads, but in the real world rocket lake is a piece of dogshit under the boot of Alder/Raptor and Zen4 still loses in 90% of applications. Goes to show my point that Intel's decision to cut AVX-512 support did not affect their product's competitiveness at all, not just for intel "sheep" but for consumers in general.

90% was optimistic but even so I was including games in my definition of applications, with 60-70% wins in productivity and 90% in games, let's call it 70-80%. Even 70-80% doesn't mean that Zen4 loses by a wide margin, it just means that overall it is a less performance product even if by a couple percentage points on average.

My entire point was responding to a graph which showed Raptor lake AVX-512 performance losing by an order of magnitude to Zen4 and raptor lake:

I only intended to show how despite these massive performance differences, in the vast majority of situations they are irrelevant, which they are. There are 0 benchmarks in any of the links I posted which have a similar performance lead shown in these graphs.

We enjoy having you here. But we have 2 rules you must abide by. Attack the post not the poster. And no profanity in tech. You have a 2 day vacation based on points.

Mod DAPUNISHER

 

[GTracing]

Nope. Let me rephrase.

AVX-512 & AVX10 are "mostly" source compatible, but NOT binary compatible. Developers now have to make builds for 2 different targets. Thats extremely bad for AMD considering their anemic client market share.

Do you have a source on that? My understanding is that AVX10/512 CPUs are completely compatible with AVX512 code. And AVX10.1 just adds a new unified feature flag, so AVX10.1 code is compatible with AVX512 CPUs other than that one caveat.

 

[Joe NYC]

Nope. Let me rephrase.

AVX-512 & AVX10 are "mostly" source compatible, but NOT binary compatible. Developers now have to make builds for 2 different targets. Thats extremely bad for AMD considering their anemic client market share.

Most of the Intel client market share is in notebooks, and I don't think Panther Lake is getting this AVX10. From the document above, it starts from Nova Lake (desktop) in 2026. Mobile - TBD.

In the meantime, AMD has been shipping AVX-512 since Zen 4 in 2022 desktop and mobile since 2023.

With AMD and Intel converging their x86 implementations, I don't see Intel introducing something that is not compatible with AVX-512.

 

[Schmide]

Nope. Let me rephrase.

AVX-512 & AVX10 are "mostly" source compatible, but NOT binary compatible. Developers now have to make builds for 2 different targets. Thats extremely bad for AMD considering their anemic client market share.

You're just trying to twist it into a smear and it won't work. AMD has implemented almost all of AVX512 (minus xeon phi extensions and FP16*) If you think intel's lesser processors will suddenly have their subset of functionality overshadow AMD's full functionality you are solely mistaken.

and BTW they are binary compatible as long as FP16 isn't used. No programmer is going to use an extension used by only one processor thus pigeonholing themselves.

*BF16 seems to be accepted as the standard as every implementation supports it. Sapphire rapids is the only processor that support FP16.

 

[Win2012R2]

Thats extremely bad for AMD considering their anemic client market share.

Not at all since AVX10 hardware isn't even shipping and first good variant 10.2 is only due 2026, if it even sells to have any meaningful presence - it also brings very little new good stuff worth creating 2nd path for even if it's relatively easy, very few people in the first place do AVX512, it's going to be come the lowest common denominator for a very long time.

The main driver for AVX512 is in servers anyway, if Turin gets real volume next year then it should dominate that space for a long time.

Intel APX will be (hopefully) a lot more successful as it brings real new meaty stuff to the table.

I am glad they killed 128b vector path so we get minimum 256b at least

I reckon they will kill 256b also and do double pumping AMD styley. The whole AVX10 should be scrapped in lieu of whatever AMD and Intel agree on going forward in that new advisory group of theirs.

 

[SiliconFly]

.

 

[Win2012R2]

Intel client market share is at around ~80%.

And AVX512 share of that market is ... 0%!

And you might be surprised but in that 20% of early AMD buyers who went Zen4/5 it might be very high concentration of people who actually run software that can benefit from AVX512, where as 80% of the market Intel got won't know, won't care.

Intel don't dictate jack now - people stayed away from AVX512 every since it was out because of their junk implementation and stupid market segmentation, AMD broke the dam and people who write AVX512 now most likely using it on AMD hardware in the first place.

The only thing that will adopt AVX10 relatively quickly are JITs, but same benefit will work for AVX512 which they will also support

Once Intel clients have AVX10, no developer is gonna spend additional time maintaining two different branches just to support AMD's AVX-512.

You right there - said developers would do what they've always done in such cases - use lowest common denominator which is AVX 512, luckily they are so close that's not even a real compromise, just stick with AVX512 and you are good. That's how it always worked and that's how it will again - with exceptional cases where Intel funds development to support AVX10, like they do in some JITs.

 

[OneEng2]

Do you have a source on that? My understanding is that AVX10/512 CPUs are completely compatible with AVX512 code. And AVX10.1 just adds a new unified feature flag, so AVX10.1 code is compatible with AVX512 CPUs other than that one caveat.

That is my understanding as well. In fact, the Intel paper I quoted from literally says this.

You're just trying to twist it into a smear and it won't work. AMD has implemented almost all of AVX512 (minus xeon phi extensions and FP16*) If you think intel's lesser processors will suddenly have their subset of functionality overshadow AMD's full functionality you are solely mistaken.

and BTW they are binary compatible as long as FP16 isn't used. No programmer is going to use an extension used by only one processor thus pigeonholing themselves.

*BF16 seems to be accepted as the standard as every implementation supports it. Sapphire rapids is the only processor that support FP16.

This!

Not to mention that if this were true, Intel's own legacy processors would be incompatible as well. With few unique extensions, AVX10 will likely have few applications that utilize these instructions, and even if they do, I don't see something like FP16 giving any serious performance boost over instructions already existing in AVX512.

It was my understanding however, that AVX10 currently attempts to implement 128b implementation of the 512b registers (quad pumped I guess?) in addition to supporting double pumped 256b and fully pipelined 512b. I think that the 128b may require a recompile? Anyway, the article over at ChipsAndCheese seems to think that Intel should seriously rethink this path.

Even if they don't, I think it will be rare to see the unique extensions used in any general purpose consumer app. It just doesn't make sense.

 

[SiliconFly]

.