Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[soresu]

Might I say the government consistently giving a win for Intel vs outsiders have to do something with it?

No, Apple's profit margins have everything to do with it

Govmt using Apple only hardware wouldn't change a thing, not to mention that Apple doesn't even make the kind of products govmt needs outside of client use, and are extremely unlikely to ever sell such a product (using inside their own company structure is a different matter).

The basic reality here is that Apple won't ever try to match the lower price range of the x86 market, and very likely Qualcomm SDX won't either.

 

[naukkis]

Try that again in FP/SIMD and M4 dies a horrible death 🤣

4x 128 bit vs 4x 512 bit isn't even a competition.

SME improves some specific things for them in SIMD - but not everything.

Apple has unified memory with GPU and well vectorizable code is practical to execute in GPU. Apple designs wins big time this competition.

 

[MS_AT]

That's why the saner side for x86 is the Intel E core team ignoring AVX-512 and straight up doubling the number of vector units like ARM has been doing.

But AVX512 is nice [I mean all the new instructions it introduced, masking things like that], the problem is the base width, that's why the most sane implementation was Zen4 [actually I thought they would double the double pumped units before Zen5 announcements, but then I was made aware that would be too costly on the silicon implementation side].

 

[gdansk]

Apple has unified memory with GPU

Doesn't Strix Point? How was the MALL to work without that?

 

[Gideon]

Try that again in FP/SIMD and M4 dies a horrible death 🤣

4x 128 bit vs 4x 512 bit isn't even a competition.

SME improves some specific things for them in SIMD - but not everything.

I personally couldn't care less. I'd much rather have zen 5 with 4x 128bit (Zen 1 style FPU) and apple like IPC at 4.5 Ghz rsther than the actual zen 5 even at 6 ghz ST.

It would also be way more useful for laptop and non-HPC server loads.

 

[soresu]

You do know FP is much easier to boost right?

#1. You are really trivialising semiconductor design here - nothing in this industry is "easy" at such an insanely fine process pitch, however it might appear looking at a µArch flow chart.

#2. Apple Ax/Mx cores have been at 4x 128 bit SIMD for significantly longer than Cortex X1-X4 was.

Just because they can add more doesn't mean that they will, especially as it will come at the cost of a significantly larger core and power draw too, which eats into their profit margins and battery running times.

They likely won't go that hard, because the vast majority of client use doesn't benefit from it.

The DCC segment of the market will take all the SIMD you can throw at it, but the vast majority of Apple users are not DCC professionals, and they certainly aren't HPC people.

The thing you neglect to realise is that Intel and AMD design their cores for both client and server/datacenter use cases.

Apple just design a generalist client use core that targets the lowest common denominator of their market, with AMX/SME added to future proof the core for AI/ML software.

 

[soresu]

Apple has unified memory with GPU and well vectorizable code is practical to execute in GPU

I've heard exactly the same argument about AMD FSA/HSA years back and it never amounted to anything.

It's still just GPGPU compute when you ignore the fluff.

 

[DavidC1]

#1. You are really trivialising semiconductor design here - nothing in this industry is "easy" at such an insanely fine process pitch, however it might appear looking at a µArch flow chart.

That is for sure. But then it has been said to be difficult 40 years ago.

All the more so they should make saner decisions, not a 5.7GHz equal to a 4.5GHz one.

The thing you neglect to realise is that Intel and AMD design their cores for both client and server/datacenter use cases.

The gains might not be so extreme, but it's still extremely large when comparing with latest Cortex cores too.

Again the ideology is at fault, which is clockspeed is king. Netburst never died, it got out of the mental hospital, received counseling and implemented it's ideas in a more subtle way.

Client vs server isn't the problem as much as Client Desktop vs Client Mobile and Server. Mobile efficiency translates well into server.

 

[SarahKerrigan]

The thing you neglect to realise is that Intel and AMD design their cores for both client and server/datacenter use cases.

Most server volume is enterprise/cloud stuff that doesn't give half a crap about SIMD or even FP.

 

[Mahboi]

Suggs likes his FPU/SIMD thicc and fast.

Suggs to have most programs use INT mainly.
(yes I'm 12 and no I'm not over his name)

 

[DavidC1]

Most server volume is enterprise/cloud stuff that doesn't give half a crap about SIMD or even FP.

Most client has stopped too. 2-cycle AVX-512 ala Zen 4 is enough, you got the benefit of the instructions anyway. Expanding from this point is where too many resources are being spent on it.

AVX-512 issues are of course stemming from the ISA coming from a deeply internally troubled company. It started from the days where they were seemingly ok and decided to stretch too far and address GPUs at the same time. Then the troubles mounted and they were stuck with AVX-512.

 

[exquisitechar]

Most server volume is enterprise/cloud stuff that doesn't give half a crap about SIMD or even FP.

Indeed. I don’t know what market they were aiming for with Zen5’s FPU. Why not allocate resources elsewhere?

 

[StefanR5R]

x64 side is penalized with only 16 GPRs, so it will spill to cache more often vs 32 GPRs on ARM

Before they spill to cache, renaming within the register file happens. The integer register file size of Zen 3/4/5 is 192/224/240.

 

[SarahKerrigan]

Indeed. I don’t know what market they were aiming for with Zen5’s FPU. Why not allocate resources elsewhere?

I mean, I'm sure the fluid-dynamics/seismic-analysis/weather-prediction kids will really like it, but that market isn't huge and a large portion of it is served by accelerators anyway.

That leaves... uh... technical-computing workstations, I guess?

FWIW I continue to think Zen5 is mostly a good gen, but it's certainly an oddly-balanced one.

 

[igor_kavinski]

All the more so they should make saner decisions, not a 5.7GHz equal to a 4.5GHz one.

OK, let's suppose AMD did do the dumb thing. What choice do consumers have now? Switch to Apple ecosystem, pay through their noses and lose access to their x86 software library? Or do they rewrite/recompile their software for Apple, assuming they run only open source software? The general public doesn't use FOSS. Businesses are more dependent on commercial x86 software than FOSS. Should we begin the great Apple silicon transition now, porting everything to Apple Silicon and sacrifice all manner of freedom? Apple's lead in ST benchmarks or perf/w means diddlysquat for the average user. x86 matters and 5.7 GHz Zen 5 is the current king for them.

 

[soresu]

Again the ideology is at fault, which is clockspeed is king. Netburst never died, it got out of the mental hospital, received counseling and implemented it's ideas in a more subtle way.

#1. It's not ideology - it's design strategy.

#2. Clockspeed is not king - that's an SKU attribute.

The problem is running SKUs at the far end of the µArch/fab node bell curve where efficiency has gone to balls.

#3. AMD are not competing with Apple (right now anyway), they are competing with Intel.

They will continue to do so until Apple is a server/datacenter chip provider.

I don't see that happening ever.

 

[DavidC1]

#1. It's not ideology - it's design strategy.

#2. Clockspeed is not king - that's an SKU attribute.

It IS ideology, because the top SKU needing 5.7-6GHz isn't achievable without it. The change needed is reducing clocks where it isn't exotic cooling domain(or failures like 13/14th gen).

They gave up on Tejas because it needed a mere 150W. What advanced was heatsink designs and size of it. Today's HSF makes Tejas HSF look like a baby brother.

#3. AMD are not competing with Apple (right now anyway), they are competing with Intel.

Right. Which is why MS is forcing all vendors to throw their weight at Qualcomm. Or why Lunarlake exists.

There is ARM too remember? The top of the line x86 cores, one that hasn't even arrived yet on the scene is barely outperforming Cortex X3. Sure, x86 clocks higher but what was the excuse before? That ARM could never reach x86 level performance per clock?

Who enabled the modern ARM push? Single handedly Apple. Who's still pushing ARM vendors to do better? Again Apple. Don't let your hatred of the fruity company bias you to make wrong conclusions(same at @igor_kavinski). I don't like their system either but their cores are top notch and by no small degree.

You can like their ideas without liking them. Opposite is true too.

 

[igor_kavinski]

I don't like their system either but their cores are top notch and by no small degree.

You can like their ideas without liking them. Opposite is true too.

I don't disagree there. But Apple having the best design doesn't do the rest of us any good. Adopting their best design will not make our life easier. It will force us to make a lot of compromises which makes the comparisons of Zen 5 with M4 plain stupid. Now if it was SD Elite X vs. Zen 5, at least the former TRIES to maintain compatibility. There's a path towards a future there without cutting off all ties to the present and the past.

 

[desrever]

Apple can optimize their whole stack from the OS level down, including their browser. And for some benchmarks, they can optimize the entire benchmark too. There just won't be any other CPU that beat them at performance/watt simply due to this. Nobody really knows how much black magic optimization is happening for Apple devices running Apple silicon. Here is Apple silicon on Linux:


M2 gets absolutely trashed by Zen 4. M4 vs Strix would probably be similar. Of course optimizations in Linux for isn't perfect so there is gains to be had under Linux for Apple silicon but my point is the hardware isn't the whole thing. The Software matters, probably way more than people give it credit.

 

[FlameTail]

Integer is that much HARDER than FP to improve. So things such as increased core-to-core latency impacts Int much greater than on FP.

Could you elaborate on why that is the case?

Also, this is another indication to me both P core teams(Intel and AMD) are struggling.

Mhmm...P-core teams everywhere seem to be struggling...

 

[SarahKerrigan]

Apple can optimize their whole stack from the OS level down, including their browser. And for some benchmarks, they can optimize the entire benchmark too. There just won't be any other CPU that beat them at performance/watt simply due to this. Nobody really knows how much black magic optimization is happening for Apple devices running Apple silicon. Here is Apple silicon on Linux:

View attachment 104180
M2 gets absolutely trashed by Zen 4. M4 vs Strix would probably be similar. Of course optimizations in Linux for isn't perfect so there is gains to be had under Linux for Apple silicon but my point is the hardware isn't the whole thing. The Software matters, probably way more than people give it credit.

This is a dubious argument. SPEC cannot be made magically faster by a OS.

 

[DavidC1]

I don't disagree there. But Apple having the best design doesn't do the rest of us any good. Adopting their best design will not make our life easier. It will force us to make a lot of compromises which makes the comparisons of Zen 5 with M4 plain stupid. Now if it was SD Elite X vs. Zen 5, at least the former TRIES to maintain compatibility. There's a path towards a future there without cutting off all ties to the present and the past.

Intel/AMD needs to do better.

I don't know if they did a dumb thing. Maybe they did not. But they need to change their IDEOLOGY. Chasing clocks again.

-Itanium failed regardless of the design or how good it was because the ideology behind needing a perfect compiler optimizing for a completely different ISA was fundamentally flawed.
-Netburst failed because they believed architectural tricks would compensate for the ways needed to get to the absurd clocks. And they didn't even meet the clock goals anyway!
-AVX-512 is also a failure because it's based on the Old Intel ideology where they can make general purpose CPUs fast enough to stem the tide of GPUs forever. Now they can't abandon it, and they can't add it efficiently either. It's too big and bulky. Of course the endless internal struggles as if the company was a microcosm of the whole of US makes matters much worse. I wouldn't be surprised if the E core team stays away from AVX10.1/10.2 as long as possible.

 

[FlameTail]

Mind sharing some of the stuff you're on?

Spoiler

 

[SarahKerrigan]

-Itanium failed regardless of the design or how good it was because the ideology behind needing a perfect compiler optimizing for a completely different ISA was flawed.

To the extent it failed at all (which it didn't, really, but it obviously isn't with us anymore) - there are reasons way beyond that; every gen late, most of them missing clock targets, IBM having damn near miraculous execution in the early 2000s, Intel/HP feuding, general decline of the RISC/UNIX market after K8, perception by potential Itanium OEMs that HP was structurally privileged in IPF, etc, etc, etc, etc

 

[Geddagod]

Spoiler

View attachment 104181

Valid avatar meme
One of my favorite shows