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

Page 67 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Mopetar

Diamond Member
Jan 31, 2011
8,141
6,838
136
Again, it's an Apple-class core.
That's really what you need to know.
Very simple.

I'm taking it to means that it's a much wider design. Depending on how much wider they make it over existing chips and how well they can keep it fed, 20% IPC could be quite conservative.

If logic is the only thing that's scaling well, then logic dictates that there should be an increase in logic. It's only logical after all.
 

adroc_thurston

Diamond Member
Jul 2, 2023
3,647
5,270
96
I'm taking it to means that it's a much wider design
Yes, that's literally where the entire industry is heading.
Things like Cortex-X4 are just insane in places (10 decode, 10 issue, 8 ALUs, three branch ports, jesus).
If logic is the only thing that's scaling well, then logic dictates that there should be an increase in logic. It's only logical after all.
Bingo, we're entering prime dark silicon spam era for CPUs.
 
  • Like
Reactions: Tlh97 and moinmoin

Mopetar

Diamond Member
Jan 31, 2011
8,141
6,838
136
I don't think we'll see much dark silicon. In the past the newer nodes meant you could cram a lot more in without having to really optimize around space. Wasted space on the newest nodes is almost twice as expensive as the same area on an older node.

Newer nodes having almost the same exponential increase in cost as they offer in transistor density suggests that designs like Zen 4c where the design is made as compact as possible will also become more of a focus.

AMD already seems to have a solution in place to deal with this. Just move more of the cache or the IO to other dies on other nodes where the cost isn't nearly as high. Stacked silicon is going to be more important than ever and I think we're still in the infancy of it.
 

adroc_thurston

Diamond Member
Jul 2, 2023
3,647
5,270
96
RDOA3 is still too fresh for me to fully trust discord slideware leakers appreciators again.
Happens, my child.
You can still chill N31 at 3.5GHz and enjoy the intended(tm) performance!
Or you can piss off back to /g/, whatever the hell you fancy.
I don't think we'll see much dark silicon.
Zen4 already sees plenty in the FP area so that's inevitable.
like Zen 4c where the design is made as compact as possible will also become more of a focus.
It's just a different core floorplan for a different fmax.
It clocks like Zen1.
Wasted space on the newest nodes is almost twice as expensive as the same area on an older node.
Yea but it's cheap logic, so why not?
 

Abwx

Lifer
Apr 2, 2011
11,591
4,408
136
Lol, what a deja vu from the Zen 4 rumors: "Zen 3 was 20% - now they got the funding, the 5nm doubles everything, the delay between releases gave them such opportunity!!1!"

Thinking like that brought us current Zen 4.

AMD is all about making economical stuff - they will still need to scale it to 4c APUs based on Zen 5. So those uber-wide cores don't really fit.

The improvement was more than 20%, it s just that the leakers didnt know what was the breakdown between frequency uplift and IPC and went wild in their speculations.

In ST the improvement was close to 30% overall.
 

yuri69

Senior member
Jul 16, 2013
554
997
136
The improvement was more than 20%, it s just that the leakers didnt know what was the breakdown between frequency uplift and IPC and went wild in their speculations.

In ST the improvement was close to 30% overall.
Some of the wild leaks:
  • IPC gains over 25%, a total performance gain of 40% - C&C
  • total IPC improvement is a staggering 30% - About 4% of this is minor core optimizations and the rest is entirely the result of the IF and memory overhaul - "AMD eng" Reddit post
  • above 20% IPC increase (heard of increase NOTABLY higher) - MLID the clown May 2021
  • 15-24% IPC increase / 28-37% ST perf - MLID the clown May 2022
The ST improvement varies a lot. 30% is definitely on the higher side tho.
 

Exist50

Platinum Member
Aug 18, 2016
2,452
3,103
136
The improvement was more than 20%, it s just that the leakers didnt know what was the breakdown between frequency uplift and IPC and went wild in their speculations.
They got everything wrong. That the end result was still good says nothing about the quality of the leaks. Especially the ones claiming even higher gains. And of course, they were all oh so confident about it. This hype train happens every gen.
 
  • Like
Reactions: H433x0n

Abwx

Lifer
Apr 2, 2011
11,591
4,408
136
Some of the wild leaks:
  • IPC gains over 25%, a total performance gain of 40% - C&C
  • total IPC improvement is a staggering 30% - About 4% of this is minor core optimizations and the rest is entirely the result of the IF and memory overhaul - "AMD eng" Reddit post
  • above 20% IPC increase (heard of increase NOTABLY higher) - MLID the clown May 2021
  • 15-24% IPC increase / 28-37% ST perf - MLID the clown May 2022
The ST improvement varies a lot. 30% is definitely on the higher side tho.

Perf improvement was 37% in MT at Computerbase, so the 40% figure was not far from the final result, whatever if the guy said more IPc and less frequency, i pointed that they didnt knew the breakdown between those two parameters.

As for the ST i said close to 30%, so you just bending my sayings, there s not only Cinebench wich btw provided 44% in MT...

They got everything wrong. That the end result was still good says nothing about the quality of the leaks. Especially the ones claiming even higher gains. And of course, they were all oh so confident about it. This hype train happens every gen.

What i said above is prove that they werent that wrong overall, better to look at numbers than to rely on counter hype...

 
  • Like
Reactions: Tlh97

Exist50

Platinum Member
Aug 18, 2016
2,452
3,103
136
Perf improvement was 37% in MT at Computerbase, so the 40% figure was not far from the final result
The claim was 40% ST, not MT. And very specifically >25% IPC gains. That article even claims 5GHz all core. None of these happened.
whatever if the guy said more IPc and less frequency, i pointed that they didnt knew the breakdown between those two parameters
Those leakers all claimed to know both. They didn't.
What i said above is prove that they werent that wrong overall
They were wrong, drastically so. Your attempt to defend them by insisting that one random number they threw out happens to align with a completely different result from real silicon is just absurd. But I'm sure that just as always, when real silicon comes out, the goalposts will move to something else.
 
  • Like
Reactions: Tlh97 and yuri69

Abwx

Lifer
Apr 2, 2011
11,591
4,408
136
The claim was 40% ST, not MT. And very specifically >25% IPC gains. That article even claims 5GHz all core. None of these happened.

Those leakers all claimed to know both. They didn't.

They were wrong, drastically so. Your attempt to defend them by insisting that one random number they threw out happens to align with a completely different result from real silicon is just absurd. But I'm sure that just as always, when real silicon comes out, the goalposts will move to something else.

In games it does roughly 5GHz all cores, this was indeed demonstrated by AMD well before launch, beside it reach this all core frequency in various loads even if it doesnt do so in CB eventually but it s close in this bench at 4.9 or so, not a big margin error.

Beside do not pump up the numbers, MLID, wich is known for his wild speculations, said 28-37% ST and even more in MT (44% in CB at Computerbase), at a time when even AMD thought that they could hit the 6GHz mark, so he s not that off considering that AMD was tight lipped and gave absolutely no tips.

Here MLID statements :


And here the statement by Red Gaming Tech, wich are more accurate than MLID and is spot on with 20-30% better ST ,wich is right, and 30-45% better MT, wich is also very accurate, so at the end it s just you that is exagerating the hype train speed.

 
  • Like
Reactions: Tlh97 and Markfw

Exist50

Platinum Member
Aug 18, 2016
2,452
3,103
136
Beside do not pump up the numbers, MLID, wich is known for his wild speculations, said 28-37% ST and even more in MT
So, you're just going to ignore every other claim he made? Or that the other leakers mentioned made in the runup? @yuri69 linked several.

At some point this has to be trolling. No reasonable person can defend those predictions.

But I guess this is just part of the pattern. Someone always has to justify why this time the hype train is totally legit.
 

gdansk

Diamond Member
Feb 8, 2011
3,108
4,823
136
I think the point is that people were generally disappointed when Zen 4 was first given some performance estimates by Lisa Su. The IPC was given as what 8%? Why would that disappoint people so much? Because it is below industry trends, sure, but they compensated for that with clock rate. People had expectations of more IPC from rumors and in some cases expecting there were would be more cores.

Not saying Zen 4 is a dud just that expectations were once again higher than reality.
 
  • Like
Reactions: Tlh97

Abwx

Lifer
Apr 2, 2011
11,591
4,408
136
I think the point is that people were generally disappointed when Zen 4 was first given some performance estimates by Lisa Su. The IPC was given as what 8%? Why would that disappoint people so much? Because it is below industry trends, sure, but they compensated for that with clock rate. People had expectations of more IPC from rumors and in some cases expecting there were would be more cores.

Not saying Zen 4 is a dud just that expectations were once again higher than reality.

They sandbagged the IPC improvement, they said 8% in an early slide and then at launch they stated 13%, wich is true in MT, in ST it s somewhat less, at least in FP.

So, you're just going to ignore every other claim he made? Or that the other leakers mentioned made in the runup? @yuri69 linked several.

At some point this has to be trolling. No reasonable person can defend those predictions.

But I guess this is just part of the pattern. Someone always has to justify why this time the hype train is totally legit.

Red Gaming Tech was accurate, i posted his claims and final results, you cant deny that he was spot on.

MLID not so but he wasnt very far from the real numbers, i posted what he claimed and the comparison with actual numbers as well, he was at least right for MT.

Now these were leaks 6 months before launch, and averaging the two sites numbers we had a good idea of the perfs to come since RGT was even below expectations for MT.

So ultimately there was not that much hype, the only thing that was missed was the ST perf by MLID, but still, his lower range was at 27% and the CPU clocked 4-5% lower than expected by AMD.
 
  • Like
Reactions: Joe NYC

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
26,258
15,388
136
So, you're just going to ignore every other claim he made? Or that the other leakers mentioned made in the runup? @yuri69 linked several.

At some point this has to be trolling. No reasonable person can defend those predictions.

But I guess this is just part of the pattern. Someone always has to justify why this time the hype train is totally legit.
Your view of how Zen 4/Genoa turned out certainly does not equal mine, or many others. The big difference is that full 100% utilization of Zen 4 cores blows away anything Intel has. No sense in me linking benchmarks, as you would just find a way to twist things to Intel. But I have have log files from WCG and primegrid that shows Genoa/Zen 4 totally dominating anything Intel.

This being the case, Zen 5 could follow the same line, as "common benchmarks" do not always tell the tale. And benchmarks using watercooling and LN2 do not impress me, as they have no real bearing in reality.

Lets talk when Zen 5 comes out. Or if you want, find some real (as in air cooled and stock ) benchmarks of SR under full 100% sustained load, and I will run the same to prove my point.

Edit: I already posted benchmarks and there is a video that declares Genoa the world record holder in many benchmarks. I have 2 9554 64 core and a 9654 96 core to run them.

Edit 2: Also power usage, SR can not touch Genoa in perf/watt.
 
Last edited:
  • Like
Reactions: Joe NYC

Abwx

Lifer
Apr 2, 2011
11,591
4,408
136
Your view of how Zen 4/Genoa turned out certainly does not equal mine, or many others. The big difference is that full 100% utilization of Zen 4 cores blows away anything Intel has. No sense in me linking benchmarks, as you would just find a way to twist things to Intel. But I have have log files from WCG and primegrid that shows Genoa/Zen 4 totally dominating anything Intel.

This being the case, Zen 5 could follow the same line, as "common benchmarks" do not always tell the tale. And benchmarks using watercooling and LN2 do not impress me, as they have no real bearing in reality.

Lets talk when Zen 5 comes out. Or if you want, find some real (as in air cooled and stock ) benchmarks of SR under full 100% sustained load, and I will run the same to prove my point.

Edit: I already posted benchmarks and there is a video that declares Genoa the world record holder in many benchmarks. I have 2 9554 64 core and a 9654 96 core to run them.

If we look at the benchmarks at Computerbase what is impressive is not the FP improvement wich is 12/13% for CB R20/R15, it s the 12/15% in 7 Zip and Handbrake, and particularly 7 Zip wich is a good metric to estimate perfs in servers workloads.

Integer related IPC is much more difficult to improve than FP and here everybody seems to think that CB, an Intel optimised bench btw, is the end of all.

As an exemple of Integer and FP IPC evolution if we compare a Core 2 to Haswell the improvement in Cinebench is roughly 60% while in 7 Zip it s a meager 12%...
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
26,258
15,388
136
If we look at the benchmarks at Computerbase what is impressive is not the FP improvement wich is 12/13% for CB R20/R15, it s the 12/15% in 7 Zip and Handbrake, and particularly 7 Zip wich is a good metric to estimate perfs in servers workloads.

Integer related IPC is much more difficult to improve than FP and here everybody seems to think that CB, an Intel optimised bench btw, is the end of all.

As an exemple of Integer and FP IPC evolution if we compare a Core 2 to Haswell the improvement in Cinebench is roughly 60% while in 7 Zip it s a meager 12%...
I am not sure what you are saying, but cinebench is NOT a real server load, but it does show performance in MT with 100% utilization. I think I am clear in saying Genoa blows away any Intel server chip at stock in sustained 100% load. Oh, and on AIR cooling, no LN2 crap.

Edit: and this is what I see for Zen 5 as well, but even more so.
 

Abwx

Lifer
Apr 2, 2011
11,591
4,408
136
I am not sure what you are saying, but cinebench is NOT a real server load, but it does show performance in MT with 100% utilization. I think I am clear in saying Genoa blows away any Intel server chip at stock in sustained 100% load. Oh, and on AIR cooling, no LN2 crap.

Edit: and this is what I see for Zen 5 as well, but even more so.

I said that 7 Zip is a good metric to estimate perfs in servers workloads, obviously AMD did put a great effort in improving Integer IPC since that s what matter in the most profitable market.

Granted they have a node advantage but that wouldnt be enough if they didnt have top notch IPC in this register, and we can expect Zen 5 to improve substancially perfs in this domain, indeed Mark Papermaster once stated that efficency was the primary concern and we all know that frequency is a dead end when it comes to improving perfs and keep power under control, so the speculations that Zen 5 IPC will see a quite big uplift are founded by the efficiency limitations that arise from frequencies uplifts.
 

Exist50

Platinum Member
Aug 18, 2016
2,452
3,103
136
Your view of how Zen 4/Genoa turned out certainly does not equal mine, or many others
This is a statement of fact. Rumors claimed huge IPC increases that were nowhere close to reality. That says nothing about how good or bad the core is, but it's been a consistent trend with early Zen rumors. There's been some of that on the Intel side too, but people don't seem to run away with it like they do with AMD.
No sense in me linking benchmarks, as you would just find a way to twist things to Intel
You mean I'd call you out if you fabricate results, as is your habit? I distinctly recall you accusing people of trolling for reading your own "source" once. So yeah, spare us the farce.
Lets talk when Zen 5 comes out
That's what we're doing about Zen 4 now, a reality check on what happened to the early hype. When Zen 5 is out and the same thing happens, I'm sure there will be people trying to brush it all under the rug in preparation for the Zen 6 hype train. To what end goal, I can't even fathom.

But hey, if you believe this, put a stake in the ground. What's your expectation for Zen 5 specint/specfp IPC gain over Zen 4?