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

[DisEnchantment]

Speculate at will

 

[igor_kavinski]

Same in terms of nanoseconds, numerically they are different, naturally. Like CL30 at 6400 is equivalent to CL37 at 7800 and so on.

Excited to see how 8800 MT/s JEDEC kits will fare with Zen 5. Those cores just chew through data and want more of it!

 

[tsamolotoff]

Excited to see how 8800 MT/s JEDEC kits will fare with Zen 5. Those cores just chew through data and want more of it!

Pretty sure IO die still is limited to 80x multi, so probably they won't fare at all

 

[Det0x]

WCCFTech have made a "news" post again about results shared in this thread

AMD Ryzen 9 9950X CPU Tested With Unlimited PPT Mode: 320W Power, Over 5.5 GHz Clocks Across All Cores, 40% Faster Vs 14900K

AMD's Ryzen 9 9950X CPU has been tested with a 5.5 GHz all core overclock and again, using an unlimited power profile pushing 320W power.

wccftech.com

But one of the comments made me lol atleast 🤣

BREAKING:
If you look at 13th gen Raptor Lake architecture very closely, you will notice they used B-Cores instead of P-Cores.
Which explains the slow degradation over time, confusion when put under load leading to frequent crashes, and word salad log files.

 

[Glo.]

Now that is a CPU with bad memory.

 

[PJVol]

Well, if you don't believe me you welcome to do what ppl at i2hard (as well as derbauer) did, they actually have multiple CPUs and delidded them. Same stuff happens to intel cpus, sometimes the difference is even greater, like 20C or so.

I'm afraid I'm missing the point. Didn't you talk about the difference in core t° in the same CCD?

 

[Goop_reformed]

WCCFTech have made a "news" post again about results shared in this thread

AMD Ryzen 9 9950X CPU Tested With Unlimited PPT Mode: 320W Power, Over 5.5 GHz Clocks Across All Cores, 40% Faster Vs 14900K

AMD's Ryzen 9 9950X CPU has been tested with a 5.5 GHz all core overclock and again, using an unlimited power profile pushing 320W power.

wccftech.com

But one of the comments made me lol atleast 🤣

Lmao I love this meme

 

[tsamolotoff]

I'm afraid I'm missing the point. Didn't you talk about the difference in core t° in the same CCD?

Yes, this difference disappears if CPU is delidded and LM is used instead of whatever AMD puts under the lid or if you DD cool it.

 

[PJVol]

But one of the comments made me lol atleast

I can only guess what P-cores look like...

 

[PJVol]

Yes, this difference disappears if CPU is delidded and LM is used instead of whatever AMD puts under the lid or if you DD cool it.

It'd be curious to see screenshots or links at the very least?

 

[tsamolotoff]

I'd curious to see screenshots or links at the very least?

It's been two years, can't find them quickly, but you can check out Roman's video, this kind of improvement would be impossible if solder was great (as it was with zen3 and below, check his delidding videos for zen3, it was like 5C of improvement in total instead of 20C), now it is intel grade, unfortunately.

 

[PJVol]

It's been two years, can't find them quickly, but you can check out Roman's video

Didn't see per core temps there. Too lazy to dig through his videos, sorry, but looking forward of seeing the data.

@igor_kavinski
I told you...

So far, he has shown 60W,90W,120W,160W, 230W PPT results, and now, we have two new entries, one with a 5.5 GHz overclock at 253W and one with unlimited PPT.

 

[Geddagod]

This graph is not exactly super optimistic for Zen 5 perf/watt...
Edit: sorry, I should give credit... I forgot to link the post I got it from. Here it is.

 

[itsmydamnation]

View attachment 103000
This graph is not exactly super optimistic for Zen 5 perf/watt...
Edit: sorry, I should give credit... I forgot to link the post I got it from. Here it is.

Unless the low tdp has the core at vmin I find that hard the believe. Very unlikely that you have better perf watt curve overall but not at the low end.

It would also make zero sense for the main target market....... Server

 

[RnR_au]

Unless the low tdp has the core at vmin I find that hard the believe. Very unlikely that you have better perf watt curve overall but not at the low end.

It would also make zero sense for the main target market....... Server

Wasn't there talk about now that AMD has Dense available that the non-Dense cores can be go further towards 'full fat' rather than being restrained according to AMD's standard Zen design philosophy? Now if you want efficient server cpus you go Dense. If you want high speed server, go non-Dense. Not a good solution for those on the desktop that likes to run 16 cores in 125W eco mode, but it is what it is.

 

[Geddagod]

Very unlikely that you have better perf watt curve overall but not at the low end.

reminds me a bit of this graph....

Chips and cheese rocket lake efficiency article

It would also make zero sense for the main target market....... Server

Ye I agree with that, but new archs generally have always had worse perf/watt uplifts at the lower end than the higher end, simply from the fact that there is more stuff to drive. Newer nodes generally followed the opposite pattern- perf/watt uplifts are better at the high end than the low end.
Look at this graph that shows Zen 3 vs Zen 2 CCX power, there are better perf/watt uplifts near the end of the graph rather than the start:

 

[Abwx]

View attachment 103000
This graph is not exactly super optimistic for Zen 5 perf/watt...
Edit: sorry, I should give credit... I forgot to link the post I got it from. Here it is.

Dunno what you are seeing, at same core count Zen 5 provide 80% better perf/Watt at isoperf, or 28% better perf/watt at 16% better perf, that s real numbers, not impressions pulled out of randomness.

 

[ErichZannMusician]

View attachment 103000
This graph is not exactly super optimistic for Zen 5 perf/watt...
Edit: sorry, I should give credit... I forgot to link the post I got it from. Here it is.

That graph, if you look at it from the top, and work your way down to lower PPT, implies that the variability between the two CCDs in the 9950X is much larger than it is for the 7950X.

 

[StefanR5R]

80% better perf/Watt at isoperf

I was about to criticize you for picking a rather extreme point from the curves. But on the other hand, this very point is the one which AMD chose as default PPT for 7950X. (And for 7900X too…)

 

[Geddagod]

Dunno what you are seeing, at same core count Zen 5 provide 80% better perf/Watt at isoperf, or 28% better perf/watt at 16% better perf, that s real numbers, not impressions pulled out of randomness.

Lol, using iso watt numbers to get a bigger figure.
We all know how much of a mid core SNC/CYPC was, and I can pull esentially that same figure of ~80% better perf/watt iso-perf for that core using the chips and cheese graph, vs skl.
At ~3 watts per core, Zen 5 is only like 0-5% better perf/watt than Zen 4, assuming 30 watts for the IOD. 3 watts is around what I expect each Zen 5 core in a Turin-flagship system to be getting.
We only start getting ~10% better perf/watt past ~6 watts per core.
We don't go past ~15% better perf/watt until ~7.5 watts per core.
It looks like Zen 5 is only able to get better perf/watt when Zen 4 is pretty much done scaling lol.

What's funny is that at 3 watts, Zen 3 shows a 17% improvement in perf/watt, and even at 2 watts it still shows a good ~13% improvement.

What's even more worrying is the possibility that these figures are from an application that shows above average gains vs Zen 4. Afterall, in AMD's presentation, they indicated Blender would be one of the workloads that would achieve an above average IPC gain over their own self reported average IPC gain figure.
I haven't even seen this same application used to model other archs though, and these are ES samples, so take everything with a grain of salt, but it doesn't look good.

That graph, if you look at it from the top, and work your way down to lower PPT, implies that the variability between the two CCDs in the 9950X is much larger than it is for the 7950X.

Hmm?

I was about to criticize you for picking a rather extreme point from the curves. But on the other hand, this very point is the one which AMD chose as default PPT for 7950X. (And for 7900X too…)

Those skus are potentially AMD's least important segment. Actually wait, I forgot about HEDT lol.

 

[StefanR5R]

At ~3 watts per core, Zen 5 is only like 0-5% better perf/watt than Zen 4, assuming 30 watts for the IOD. 3 watts is around what I expect each Zen 5 core in a Turin-flagship system to be getting.

I think we still lack data to start extrapolating to Turin. Genoa and Raphael supposedly have differently binned CCDs, yet Turin and Granite Ridge¹ allegedly get different steppings, don't they? Although if the latter is true, the reasons and effects are uncertain.

________
¹) of which we only have ES co-tuned data for now

 

[Abwx]

I was about to criticize you for picking a rather extreme point from the curves. But on the other hand, this very point is the one which AMD chose as default PPT for 7950X. (And for 7900X too…)

The point for server FI is to pack 80% more CPUs and get 80% more throughput at same power, dunno wich other metric could be of importance for this segment.

For DT that s another matter, either you shoot for perfs + perf/watt and you ll trade some efficency for more perfs, in wich case you ll get 1.16x the perfs at 0.9x the power FI.

Or you can dump all the perf/Watt improvement to get the highest possible perfs and set say a 9950X at 230W PPT, wich would provide 20% better perfs, and perf/Watt, at same power than the 7950X.

That being said the curves for those Blender runs are truncated because the IOD use much more power than at stock, hence at the lower powers the power/frequency scaling range that looks linear is exageratly stretched, the curve should converge to a parabol much earlier.

 

[Abwx]

Lol, using iso watt numbers to get a bigger figure.
We all know how much of a mid core SNC/CYPC was, and I can pull esentially that same figure of ~80% better perf/watt iso-perf for that core using the chips and cheese graph, vs skl.
At ~3 watts per core, Zen 5 is only like 0-5% better perf/watt than Zen 4, assuming 30 watts for the IOD. 3 watts is around what I expect each Zen 5 core in a Turin-flagship system to be getting.
We only start getting ~10% better perf/watt past ~6 watts per core.
We don't go past ~15% better perf/watt until ~7.5 watts per core.
It looks like Zen 5 is only able to get better perf/watt when Zen 4 is pretty much done scaling lol.

All those numbers of yours are pure non sense.

Efficency delta between a Zen 4 core and Zen 5 core is the same at 3W than at 6W, there s no way that Zen 5 efficency could increase relatively to Zen 4 at the rate you re pretending, so it s clear that something is heavily flawed in your random methodology.

To summarize if Zen 5 is 15% more efficent at 7.5W than Zen 4 then it will be also 15% more efficent at 3W, , or else the laws of physics would mean nothing.

 

[Geddagod]

All those numbers of yours are pure non sense.

Efficency delta between a Zen 4 core and Zen 5 core is the same at 3W than at 6W, there s no way that Zen 5 efficency could increase relatively to Zen 4 at the rate you re pretending, so it s clear that something is heavily flawed in your random methodology.

To summarize if Zen 5 is 15% more efficent at 7.5W than Zen 4 then it will be also 15% more efficent at 3W, , or else the laws of physics would mean nothing.

Those numbers are pulled straight out of the graph lmao.
If you want me to write out the math, here you go:
3 watts per core x 16 cores + 30 watts IOD = ~80 watts power, looking at the graph, the % higher performance Zen 5 has over Zen 4 at that power is 0-5%.
6 watts per core is 6 x 16 + 30 watts IOD = ~130 watts power, looking at the graph, the % higher performance Zen 5 has over Zen 4 at that power is 7-12%.
If you want to be pedantic, ig I should be writing cores+cache+ring every time, but that same thing should be applied to both power figures reported from AMD (as they also report CCX power in their graphs) and also Turin estimations.

Also, no, if Zen 5 scores 15% higher at 7.5 watts, there is no guarantee it will also score 15% higher at 3 watts. Literally in the graph AMD presented for Zen 3:

They show how 8 Zen 3 cores has a lower perf/watt increase at 10 watts than they do at 35 watts vs Zen 2.
And again, the graph from C&C showing RKL vs Skl perf/watt show that that rule is not followed:

CYPC has like 25% better perf/watt at 7.5 watts per core vs SKL, but a 0% advantage at 3.5 watts per core.

 

[StefanR5R]

[...] or else the laws of physics would mean nothing.

Power scaling is surely not uniform over frontend logic, execution logic, the various buffers and caches, fabrics, and so on, is it?

 

[Abwx]

Those numbers are pulled straight out of the graph lmao.
If you want me to write out the math, here you go:
3 watts per core x 16 cores + 30 watts IOD = ~80 watts power, looking at the graph, the % higher performance Zen 5 has over Zen 4 at that power is 0-5%.
6 watts per core is 6 x 16 + 30 watts IOD = ~130 watts power, looking at the graph, the % higher performance Zen 5 has over Zen 4 at that power is 7-12%.

The IOD take 35W in those Blender runs, that s displayed on the pics, at 80W this amount to 2.81W/core, if a 7950X IOD use 20W then at 80W this will amount to 3.75W left per core.

You can see that the footprint of the IOD can result at 80W to roughly 33% more power per core for the 7950X, that s basically what you are doing, that is comparing two cores at 33% power difference and then doing wrong conclusions out of this flawed basis, FTR 33% more power allow 10% higher clocks at those frequencies.

Power scaling is surely not uniform over frontend logic, execution logic, the various buffers and caches, fabrics, and so on, is it?

Yes but the scaling will be the same for both Zen 4 and 5, why should those elements scale differently with Zen 5.?