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

[DisEnchantment]

Speculate at will

 

[DrMrLordX]

The former figure was given by a random internet poster. Oh boy...

It wasn't just a random internet poster. It was an OEM leak picked up by a few random internet posters.

 

[biostud]

So i'm still very much confused about the discrepancy between the 32% specint 2017 and the 16% figure directly from amd. Weird.

The 16% from AMD is the geomean of different benchmarks, some are higher, some lower.

 

[DisEnchantment]

The 16% from AMD is the geomean of different benchmarks, some are higher, some lower.

He can read the AMD slides, no problem.

He is just wondering how comes it is not 32% as advertised by the leakerz. It is simple, he got bamboozled..

 

[Tuna-Fish]

I wish Intel would've shrunk Raptor Lake Refresh to 20A instead of pushing it wayyyy beyond its limits.

With the release of MSI's CAMM2 mobo, I'm not sure anymore that Zen 6 will be released on AM5. Maybe on AM5+ CAMM2 mobo.

You can do AM5 + CAMM, compatible with AM5 + DIMM at the socket. If you just intend to support DDR5 CAMM, not LPDDR5X + CAMM, there is no reason to make the socket incompatible.

(edit: ) Actually, let's make that a prediction. I predict that sometime between now and 2027, AMD/partners will release a new motherboard platform that differs from AM5 in that it supports CAMM, and this will be the best platform to run Zen6 on. But, those same CPUs still work just fine, just with reduced memory clock, on AM5 boards.

 

[Goop_reformed]

He can read the AMD slides, no problem.

He is just wondering how comes it is not 32% as advertised by the leakerz. It is simple, he got bamboozled..

The 32% figure just doesn't appear out of thin air, and that's the issue.

Now that Mlid credibility has been improved, I'm inclined to believe the new core was plagued with bugs as mentioned recently.

 

[poke01]

https://twitter.com/x/status/1798652423462088816

It seems it was a last minute name change for Strix Point

 

[StefanR5R]

I was assuming that it is still 2 CCX per die like it is in zen4c. Do you have something indicating that they changed that?

There still is merely the old rumor of a 16 core complex, no public info from AMD.

I also haven't really seen any indicating that it actually has more than 12 infinity fabric links on the IO die.

AMD subsequently gave minor clarifications to the press. Renderings from AMD:

Turin classic, up to 128 cores	Turin dense, up to 192 cores

(source)

 

[blackangus]

So i'm still very much confused about the discrepancy between the 32% specint 2017 and the 16% figure directly from amd. Weird.

Have I missed it? Have specint scores been posted?
I hadn't seen any yet.
Not that I am expecting something amazing like 32% specint 2017, but saying this set of graphs have a 16% number and expecting it to be 32% when the actual benchmark specified for 32% isn't on the graph seems like the expectation is off there.

 

[inf64]

Have I missed it? Have specint scores been posted?
I hadn't seen any yet.
Not that I am expecting something amazing but....

Nope, nothing was posted except the poor breakdown of IPC on Computex.

 

[StefanR5R]

Now that Mlid credibility has been improved,

Nothing has changed of what we know about MLID's quality of sources, background checking, interpretations, and outright inventions.

 

[igor_kavinski]

outright inventions.

An essential ingredient for hype trains.

 

[leoneazzurro]

The 32% figure just doesn't appear out of thin air, and that's the issue.

Now that Mlid credibility has been improved, I'm inclined to believe the new core was plagued with bugs as mentioned recently.

32% figure was about a 1T test (not usually power or thermally limited) and it was a single data point anyway.
16% Geomean as AMD gave us was determined on all nT tests (which are power/thermally limited)
Process thermal/power improvements going from N5 to N4P are on the order of magnitude shown in AMD's nT tests.

So I am inclined to say that in standard nT applications is unlikely we'll see an average greater than the 16% advertised by AMD. Exception could be application with heavy FP load and for sure those which heavily leverage on AVX 512.
I am also inclined to think ST/low threaded applications may have an higher average improvement, as seen in the gaming tests. +32% in average, even in those cases, I think is too much.

 

[StefanR5R]

16% Geomean as AMD gave us was determined on all nT tests (which are power/thermally limited)

This slide showed iso-clock tests.¹

________
¹) Presumably, I should add. (The slide deck is, as has been pointed out by others, not free from mistakes.)

 

[eek2121]

Is there any indication it supports higher speed DDR5 than previous versions?
I'm not too optimistic on that.

The chipset is actually unchanged, so the differences will come down to board quality and BIOS improvements. Zen 4 already supports DDR5-8000, but you have to drop to 1:2 mode, which causes quite a hit.

For the 1 CCD parts, I think its pretty obvious that the X3D parts are tempting. But with the 2 CCD parts, I'm definitely gonna go with the one that does not rely on a flimsy thread prioritization routine, and is available earlier, the 9950X. I of course can't claim to know if the majority of potential buyers think like me though.

I made the mistake of going 7950X (non-3D) last time. I'm waiting for the 3D part this time around.

You can do AM5 + CAMM, compatible with AM5 + DIMM at the socket. If you just intend to support DDR5 CAMM, not LPDDR5X + CAMM, there is no reason to make the socket incompatible.

(edit: ) Actually, let's make that a prediction. I predict that sometime between now and 2027, AMD/partners will release a new motherboard platform that differs from AM5 in that it supports CAMM, and this will be the best platform to run Zen6 on. But, those same CPUs still work just fine, just with reduced memory clock, on AM5 boards.

Desktop is likely to move to LPDDR before we see non LPCAMM adoption. I will be shocked if we continue to have a split since combining the tech means lower costs and higher margins.

 

[leoneazzurro]

This slide showed iso-clock tests.¹

________
¹) Presumably, I should add. (The slide deck is, as has been pointed out by others, not free from mistakes.)

Yeah, it seems I missed that part but in any case the setup is very strange because they should also say a CCD was disabled on the 9950X to get those data, and there are other information missing as well . Also it would have been quite easy for AMD to move that average on an higher number, by including more games (Far Cry 6 seems to be one of the lowest gains looking at the comparison with the 14900K) or FP-intensive apps. It's strange because basically Zen5 seems stronger in the comparison against the 14900K than in the one against the Zen4. True that the former has no AVX512, but in some cases (gaming) has no effect or in others (CB24 supports AVX512 but AFAIK it has very small effect).

 

[JustViewing]

So i'm still very much confused about the discrepancy between the 32% specint 2017 and the 16% figure directly from amd. Weird.

My guess is, it was supposed to be a comparison between server Z4 Z5 CPUs. This somehow got translated to desktop IPC by few people.

 

[Tigerick]

So i'm still very much confused about the discrepancy between the 32% specint 2017 and the 16% figure directly from amd. Weird.

If you really want to know about SPECint and IPC, you should check my post here two months ago. I may not technical competent, but I spent time trying to calculate the figures and present in the table format to understand the difference between performance numbers and IPC. I don't have much free time, maybe you should study the table and do the technical analysis which clearly the forum members lacking...

Thanks to @Timorous for explanation.

 

[igor_kavinski]

basically Zen5 seems stronger in the comparison against the 14900K than in the one against the Zen4.

That's a smart attempt to gain marketshare especially in light of Intel's shader compilation crashes.

"Hey, miserable annoyed angry Intel customer, how about faster performance at lower wattage and heat? Interested?"

What are these users/gamers gonna do? Wait till October?

 

[Tuna-Fish]

Desktop is likely to move to LPDDR before we see non LPCAMM adoption. I will be shocked if we continue to have a split since combining the tech means lower costs and higher margins.

Desktop is moving to LPDDR6, when it comes. The question is what happens for the last two years before that. I think DDR5 CAMM2 is the likely path.

 

[Jan Olšan]

32% figure was about a 1T test (not usually power or thermally limited) and it was a single data point anyway.
.

Wasn't the whole issue that the 32% value was presented as a 1T datapoint, but (correct me if I am wrong in this understanding) the original source actually gave it as MT (across all threads) gain for the Epyc 9005 classic SKU?

I assumed what happened was that the leakers thought that Epyc 9005 classic had to have 96 cores like Genoa, so the +32/ MT specint result means it has to soemhow get +32% performance per core. However, Epyc 9005 Classic apparently goes up to 128 cores, so part of that +32 % are the extra cores, and it was a mistake to interpret it as ST gain. So wasn't that what happened? Would make a lot of sense.

 

[Goop_reformed]

Nothing has changed of what we know about MLID's quality of sources, background checking, interpretations, and outright inventions.

You're objectively wrong. His info about zen 5 has not been discredited, yet.

 

[biostud]

He can read the AMD slides, no problem.

He is just wondering how comes it is not 32% as advertised by the leakerz. It is simple, he got bamboozled..

As we can see, if you focus on only a single benchmark then it could be 35% increase in IPC, so maybe somewhere an "up to 35%" transformed to 35% increase.

 

[Goop_reformed]

As we can see, if you focus on only a single benchmark then it could be 35% increase in IPC, so maybe somewhere an "up to 35%" transformed to 35% increase.

It's just specINT is THE benchmark. Nothing burger if it's cinebench otherwise.

 

[Abwx]

It's strange because basically Zen5 seems stronger in the comparison against the 14900K than in the one against the Zen4. True that the former has no AVX512, but in some cases (gaming) has no effect or in others (CB24 supports AVX512 but AFAIK it has very small effect).

It s 21% faster than the 14900K in Cinebench 2024 according to their slide, wich seems quite possible because the 7950X is 1% faster in CB R20 + CB R23 averaged if the 14900K is set at 253W, wich is obviously the default setting used by AMD.

1.01 x 1.17 = 1.1817, so as i already pointed there s likely 2.4% all cores frequency uplift since 17% better IPC in CB wont exhaust N4P s 28% better perf/Watt than N5.

Intel Core i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks

i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks / Testmethodik / Multi-Core-Leistung / Single-Core-Leistung

www.computerbase.de

 

[leoneazzurro]

Wasn't the whole issue that the 32% value was presented as a 1T datapoint, but (correct me if I am wrong in this understanding) the original source actually gave it as MT (across all threads) gain for the Epyc 9005 classic SKU?

I assumed what happened was that the leakers thought that Epyc 9005 classic had to have 96 cores like Genoa, so the +32/ MT specint result means it has to soemhow get +32% performance per core. However, Epyc 9005 Classic apparently goes up to 128 cores, so part of that +32 % are the extra cores, and it was a mistake to interpret it as ST gain. So wasn't that what happened? Would make a lot of sense.

IIRC it was normalized for the core number so no, but I can be wrong.