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

[DisEnchantment]

Speculate at will

 

[Abwx]

Ah yes, that one subtest explains the 15% difference found by geekerwan lol.

The same geekerwan that find that Golden Cove has higher IPC than Raptor cove, that is, Intel s latest and slightly improved uarch is slower than the previous one.

That say a lot about the competence of this random guy who is in now way a reference, just some youtuber or tweetos, as accurate as your 15% that are actually 12.8% in his dubbious numbers.

Even in CB RPL doesnt manage to outmatch Zen 4 IPC by this amount, and that s the most favourable FP bench, so much for even those 12.8%

 

[adroc_thurston]

That is for the engineering sample of Zen 5. For reference I get around 740-750 on my stock 7950X, so about 21% higher. Not bad

Oooh, nice, 21% with unchanged scalar FP throughput or latencies.

 

[SarahKerrigan]

There s Povray that is included in Spec_FP, and as said it doesnt use AVX2
for AMD, so much for the clowning, i guess that you didnt even notice that this
bench was in SPEC.

FTR Zen 4 perform 16-18% better in Povray once AVX 2 is enabled for this CPU, now how much those 16-18% would amount in an average that comprise 13 benches.?..

SPEC doesn't use AVX2 for anyone, except what compilers can autovectorize, and Geekerwan wasn't using icc to build it.

 

[adroc_thurston]

SPECfp isn't a good test, don't do it.
Use the int part of the suite, it's a lot more relevant.

 

[Abwx]

SPEC doesn't use AVX2 for anyone, except what compilers can autovectorize, and Geekerwan wasn't using icc to build it.

Since SPEC use Povray it use AXV2 for this bench since Povray is implemented with AVX2 for years now, it wasnt updated to make use of AVX2 in AMD CPUs for some reasons, even if such an update would change almost nothing to the code.

That being said if we look at Computarbase ST test at same frequency Zen 4 is 2% faster than the 12900K in CB R15 while the latter is 7% faster in both CB R23 and Povray, if AVX 2 was activated for Zen 4 it would be 9-10% faster than the 12900K in the latter.

In MT IPC Zen 4 is 2% faster on an average of 10 tests including Povray.

Page 460 - Discussion - Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

Page 460 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[Geddagod]

The same geekerwan that find that Golden Cove has higher IPC than Raptor cove, that is, Intel s latest and slightly improved uarch is slower than the previous one.

That say a lot about the competence of this random guy thjat is in now way a reference, just some youtuber or tweetos, as accurate as your 15% that are actually
12.8% in his dubbious numbers.

Lmao, ig geekerwan is incomptent now. Welp.
P.S., even using Raichu's testing, removing POVray, or adding back POVray with the performance speedup for Zen 4, will still net you RPL beating Zen 4 in FP IPC, at 6000 mhz ram at least (if I can use my calculator right lol).
Same story with Anandtech

 

[SarahKerrigan]

Since SPEC use Povray it use AXV2 for this bench since Povray is implemented with AVX2 for years now, it wasnt updated to make use of AVX2 in AMD CPUs for some reasons, even if such an update would change almost nothing to the code.

SPEC has all of the assembly fastpaths for all CPUs removed. There is zero AVX2 code in SPECfp's Povray or in any other SPEC subtest.

 

[adroc_thurston]

Why are you arguing specFP, that's really not the relevant part of Spec2017.

 

[Geddagod]

Why are you arguing specFP, that's really not the relevant part of Spec2017.

It's an FP test, which has RPL having higher IPC than Zen 4. Apparently the only way Intel can win a benchmark, according to some people at least, is if Intel paid off all the reviewers, the creators of the benchmark have been paid off by Intel, or the reviewers are doddering idiots.

 

[adroc_thurston]

It's an FP test, which has RPL having higher IPC than Zen 4.

Ah, yeah.
But Intel itself declared a fatwa against specFP subtests on that slide.
If you wanna do Intel bull stuff, don't mention specFP then!

 

[Abwx]

Why are you arguing specFP, that's really not the relevant part of Spec2017.

Look here where are geekerwan numbers.?..

AMD Ryzen 7000 im Test: So schnell sind 7950X und 7700X

AMD Ryzen 7000 ist da. Im ausführlichen Test treten Ryzen 9 7950X und Ryzen 7 7700X mit Zen-4-Kernen gegen Ryzen 5000 und Intel Core an.

www.computerbase.de

If the difference was 12.8% that would appear in all the FP tests, namely Blender, CB R15 and R20, Povray and Corona, but only in CB R20 and in the AMD gimped Povray there s an advantage for ADL.

Beside if Intel had such a big advantage they wouldnt had been caught by SPEC org cheating the benches, lol.

 

[adroc_thurston]

Look here where are geekerwan numbers.?..

Who cares, use the SIR numbers.

 

[Gideon]

No, because it was modded in 2017 to greatly favour Intel after Zen 1 yielded much better scores than SKL, that s the most possible biaised bench, they alleged that it was due to a bug but that was plain lies, that was just $$ at play.

I wouldn't call it malice, it's probably just incompetence. Chips & Cheese made a good in-depth article about how useless it is on modern hardware:

CPU-Z’s Inadequate Benchmark

CPU-Z is a hardware information tool from a company called CPUID, not to be confused with the CPUID instruction. Besides showing basic CPU, motherboard, and memory information, CPU-Z features a bui…

chipsandcheese.com

 

[SarahKerrigan]

Look here where are geekerwan numbers.?..

AMD Ryzen 7000 im Test: So schnell sind 7950X und 7700X

AMD Ryzen 7000 ist da. Im ausführlichen Test treten Ryzen 9 7950X und Ryzen 7 7700X mit Zen-4-Kernen gegen Ryzen 5000 und Intel Core an.

www.computerbase.de

If the difference was 12.8% that would appear in all the FP tests, namely Blender, CB R15 and R20, Povray and Corona, but only in CB R20 and in the AMD gimped Povray there s an advantage for ADL.

There is no "AMD gimped Povray" in SPECfp.

I agree with @adroc_thurston, weirdly, that SPECfp is mostly useless (it is a poor proxy for literally any FP-intensive application) - but there is no magic built-in Intel advantage on it. None.

 

[uzzi38]

It's a terrible benchmark but AFAIK CPU-Z should be a much bigger uplift on Zen5. Also there are no IF/IMC changes so I'm skeptical of this post.

No CPU-z is one of the low ballers. As far as I'm aware at least anyway, and there's a good reason for it (that I won't be discussing).

That being said other claims made by the poster in that tieba link (e.g. the claim that Zen 5 will launch after HotChips) and the claims regarding IF speeds are extremely suspect, so I'm not sure it's worth looking too hard at their claims.

 

[Abwx]

I wouldn't call it malice, it's probably just incompetence. Chips & Cheese made a good in-depth article about how useless it is on modern hardware:

CPU-Z’s Inadequate Benchmark

CPU-Z is a hardware information tool from a company called CPUID, not to be confused with the CPUID instruction. Besides showing basic CPU, motherboard, and memory information, CPU-Z features a bui…

chipsandcheese.com

There is no "AMD gimped Povray" in SPECfp.

I agree with @adroc_thurston, weirdly, that SPECfp is mostly useless (it is a poor proxy for literally any FP-intensive application) - but there is no magic built-in Intel advantage on it. None.

I was talking of Povray, not of SPEC_fp as a whole, but then how do you explain that for Povray RPL is 13.5% faster than Zen 4 in the numbers provided by Geddagod.?.
You ll notice that there s nothing of the sort with Blender.

Mind you Computerbase measure 12% better ST perf in Povray for the 13900K,
that about the same number if we account for the frequency difference.

Core i9-13900K, i7-13700K & i5-13600K: Gaming-Könige im Test: Benchmarks in Apps

Intel Raptor Lake im Test: Benchmarks in Apps / So wurde getestet / Leistung in Single-Core-Lasten (klassisch)

www.computerbase.de

 

[SarahKerrigan]

i ws talking of Povray, not of SPEC_fp as a whole, but then how do you explain that for Povray RPL is 13.5% faster than Zen 4 in the numbers provided by Gedaggod.?.
You ll notice that there s nothing of the sort with Blender.

Mind you Computerbase measure 12% better ST perf in Povray for the 13900K,
that about the same number if we account for the frequency difference.

Core i9-13900K, i7-13700K & i5-13600K: Gaming-Könige im Test: Benchmarks in Apps

Intel Raptor Lake im Test: Benchmarks in Apps / So wurde getestet / Leistung in Single-Core-Lasten (klassisch)

www.computerbase.de

How do you explain it? It's a simple fact that there are no assembly fastpaths in SPEC, whatsoever, anywhere. There is no use of AVX2 assembly or intrinsics. All that's running in the Povray subtest is what your compiler spits out from the C++ input source.

Single subtests can surprise you. As an example, Power9 punches well above its weight in my testing on 631.deepsjeng_s and 605.mcf_s. Does that mean that SPEC is secretly biased for IBM on those subtests? No; it means that something about the microarchitecture is favorable for the particular code stream that GCC 7.3 emitted for Power9 when I ran the tests.

With enough profiling, the reason usually becomes clear.

 

[igor_kavinski]

Physics disagrees (and it doesn't listen to management).

Funny fact: I won a place in a Physics Talent Contest in my 12th year of education (call it whatever according to your place/country of residence).

Spoiler: Amusing OFF-TOPIC Anecdote

Here's the story:

So this Physics professor that almost everyone agreed was a major snoozefest (he could hardly raise his voice) asked one day in class: does anyone want to compete in a nationwide physics talent contest, the winners of which will represent their country in the International Physics Olympiad? Everybody looked around, waiting to see who would raise their hands. I believe no one dared so I thought, what the heck? I'm not gonna win and I'm not that into Physics either so fat chance this professor is gonna select me. I raise my hand. Students are surprised somewhat. Maybe there were jokes said out aloud. I don't remember. It's hazy and I was oblivious to everything at that moment anyway. Another student also raises his hand. Professor asks us both to meet him outside after class. I don't remember about the other guy but I remember talking to the professor after class and he seemed pleased that I had the courage to do something like this. Don't remember the rest but a week later I was in a room in the radiology wing of the largest hospital in my city, taking a Physics test! I have no idea why they chose that venue for the exam. Anyway, it was something like an hour or two hour exam with multiple choice questions and some that required short answers. They were fun questions like "If a housefly is flying inside a glass jar, does the fly's weight make a difference to the jar's weight?". You get the idea. And I was selected from my whole city, with two other guys who also happened to be from my academic institute. So off we went on a one week all expenses paid trip to a top secret location. Well, I don't know. Maybe it was not that secret but it had MP soldiers acting as guard duty. It was like a mini city of research institutes, dorms, houses for employees and two research nuclear reactors. I saw both of them on the inside on a tour. Forgive me if I posted about this before. I can have bad memory. Anyway, over the week, they tested our Physics thinking skills with various projects. I wasn't really into physics and also coz I was uncomfortable around chicks, I did not participate that much in my group that included two hot chicks. The faculty looked at me weird, like how the hell did this guy get here???? They even asked the others why I wasn't participating and I don't know they made up some weird answer that I don't remember. The faculty did not look pleased but what could they do? Send me back and admit that they messed up in the selection process? They got even more alarmed when they asked me later in a routine survey of everyone how many marks I got in my previous year's physics final exam. The look on the guy's face was priceless when I told him that I had flunked the previous year's exam and that I would take it again with this year's exam together! Oh man, I really must've contributed hell of a lot to improving their selection criteria in the future! By the way, I was number 49 out of 50 but for some reason, I became number 50 out of the 50 students there. I have no idea why they demoted me A reason would've been nice. I was too afraid and embarrassed to ask anyway, since I was an imposter there So concludes my brush with an opportunity of being a real physicist in life that I totally bungled because of two chicks!

 

[igor_kavinski]

Single subtests can surprise you.

Is there a free opensource analog of SPEC that mere mortals can use for compilation/benchmarking?

 

[Abwx]

How do you explain it? It's a simple fact that there are no assembly fastpaths in SPEC, whatsoever, anywhere. There is no use of AVX2 assembly or intrinsics. All that's running in the Povray subtest is what your compiler spits out from the C++ input source.

A member of this forum recompiled Povray to enable AVX2 for AMD and said that the score of his 7950X increased by 18% or so and it became faster than the Intel CPU.

So you are telling me that once AVX2 is removed for RPL it will still score faster than the 7950X and hence that AVX2 is useless for RPL in Povray, AVX2 or not it will score the same but the 7950X with AVX2 will score 18% better.?...

 

[SarahKerrigan]

A member of this forum recompiled Povray to enable AVX2 for AMD and said that the score of his 7950X increased by 18% or so that and it became faster than the Intel CPU.

So you are telling me that once AVX2 is removed for RPL it will still score faster than the 7950X and hence that AVX2 is useless for RPL in Povray, AVX2 or not it will score the same but the 7950X with AVX2 will score 18% better.?...

I'm not making larger conclusions. I am telling you that there is no AVX2 for Povray in SPECfp on AMD or Intel, nor is there VMX on Power. You realize that SPEC is compiled from source, right? There's no secrets around this. Every licensee has the full source code. The lack of platform-specific, vendor-specific, or ISA-specific backends is verifiable fact.

 

[igor_kavinski]

Zen 5 CPU-Z single core benchmark is supposedly 910.

August is enough time for them to somehow manage 999. I absolutely demand it and WILL declare Zen 5 a failure if they do not accede to my very reasonable demands.

Asking for a ST score of 1000 would be unreasonable. See? I'm a very fair minded person. Totally just and understanding.

And before anyone complains, in this speculative thread that often devolves into everyone throwing coins into the wishing well of infinite computing power, I am and remain the perfect picture of sane wistfulness.

 

[Abwx]

I'm not making larger conclusions. I am telling you that there is no AVX2 for Povray in SPECfp on AMD or Intel, nor is there VMX on Power. You realize that SPEC is compiled from source, right? There's no secrets around this. Every licensee has the full source code. The lack of platform-specific, vendor-specific, or ISA-specific backends is verifiable fact.

I agree but this doesnt explain the discrepancy in the tests posted by Geddagod, the only explanation of this discrepancy is that RPL gain nothing with AVX2 in Povray while the 7950X sure gain 18%, i find this contradictory, if Zen 4 gain such a percentage there s no reason that RPL wouldnt.

Now we dont know if his source has recompiled the tests such that they use AVX2 since that s not an official source.

 

[Hans Gruber]

August is enough time for them to somehow manage 999. I absolutely demand it and WILL declare Zen 5 a failure if they do not accede to my very reasonable demands.

Asking for a ST score of 1000 would be unreasonable. See? I'm a very fair minded person. Totally just and understanding.

And before anyone complains, in this speculative thread that often devolves into everyone throwing coins into the wishing well of infinite computing power, I am and remain the perfect picture of sane wistfulness.

From compiling all of the online speculation over the months. The conclusion for Zen 5 seems to be a large uplift in multicore processing power over Zen4 and as little as 10% uplift in single core performance. Not sure if that also means about 10% uplift in IPC.

 

[gdansk]

From compiling all of the online speculation over the months. The conclusion for Zen 5 seems to be a large uplift in multicore processing power over Zen4 and as little as 10% uplift in single core performance. Not sure if that also means about 10% uplift in IPC.

MT increase is typically < ST increase when TDP is unchanged
Where are you getting this