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

[DisEnchantment]

Speculate at will

 

[Philste]

Would you have a link?

Beitrag im Thema 'Instabile Raptor-Lake-CPUs: Intel identifiziert zu hohe Spannung als eine Ursache' https://www.computerbase.de/forum/t...annung-als-eine-ursache.2203820/post-29612118

 

[jdubs03]

finally this is impressive, almost better than 7945hx with 4 less cores and presumably less watt and thermals

I know it’s not exactly the right comparison (since as noted before the 8945H/S, is more similar), but the 7845HX has 12 cores.
The average MT shows as - 13738, but some recent posts show 15500-16000.
ST is 2600-2700. So some gain in ST for the HX 370 there. But MT, eh not that impressive.

Power consumption aside, 7845HX scores around 26900 in Cinebench R23 MT. Whereas the leaks from the HX 370 have that scoring 23200 (though the single core is a little better). When considering power consumption, for the R23 score, the HX 370 should be consuming a lot less power (just how much is the important part).

What I pointed out yesterday though was that in Cinebench 2024, the HX 370, scores 1525. Whereas (adding for brevity since I didn’t mention it earlier) the 7845HX scores around 1385.

So it trades blows in those two benchmarks.
I guess my question is what happened between Cinebench R23 and 2024, for the multicore scores differential between these two chips to be so different?

 

[MS_AT]

I know it’s not exactly the right comparison (since as noted before the 8945H/S, is more similar), but the 7845HX has 12 cores.
The average MT shows as - 13738, but some recent posts show 15500-16000.
ST is 2600-2700. So some gain in ST for the HX 370 there. But MT, eh not that impressive.

Power consumption aside, 7845HX scores around 26900 in Cinebench R23 MT. Whereas the leaks from the HX 370 have that scoring 23200 (though the single core is a little better). When considering power consumption, for the R23 score, the HX 370 should be consuming a lot less power (just how much is the important part).

What I pointed out yesterday though was that in Cinebench 2024, the HX 370, scores 1525. Whereas (adding for brevity since I didn’t mention it earlier) the 7845HX scores around 1385.

So it trades blows in those two benchmarks.
I guess my question is what happened between Cinebench R23 and 2024, for the multicore scores differential between these two chips to be so different?

R23 workload will fit in 7845HX cache, R24 won't I guess after what we have heard so far that R24 is putting more pressure on memory subsystem

 

[Goop_reformed]

Anything more concrete?

CapFrameX as a source

AMD themselves basically confirmed that ZEN5 is bad

My source is a VAR channel which focuses on professional applications. Ram timing might play a role but I wasn't informed of that. Performance is exactly what amd has advertised minus a few percent.

 

[gaav87]

Anything more concrete? Is it below Zen4, below Zen1? Do you mean desktop or laptop flavor?

Anyway I was wondering are decoders in Zen5 statically partitioned between SMT threads, so when SMT threads is enabled each thread ends up with 4 wide decoder and only if you disable SMT in the BIOS it might happen that both decoders will be trying to decode the same instruction stream?

I would not bother to take what he said serious half of his posts are "Is zen5 good yet?" "So is it good yet?" "I've not recovered from the 32% hangover ." "So what's the consensus here, is zen 5 good yet?"

 

[inf64]

My source is a VAR channel which focuses on professional applications. Ram timing might play a role but I wasn't informed of that. Performance is exactly what amd has advertised minus a few percent.

Or it could be that the lower TDPs mean the chips are not boosting as high as Ryzen 7000 in games. We'll see soon enough. Looking at the core itself (as the platform is the same), there is no reason for Zen 5 to show mediocre uplift in games.

 

[MS_AT]

Or it could be that the lower TDPs mean the chips are not boosting as high as Ryzen 7000 in games. We'll see soon enough. Looking at the core itself (as the platform is the same), there is no reason for Zen 5 to show mediocre uplift in games.

The rumored changes to boost L3 density? Maybe latency got worse and DRAM tuning plays greater role

 

[inf64]

The rumored changes to boost L3 density? Maybe latency got worse and DRAM tuning plays greater role

I doubt that is the case.

 

[Philste]

Or it could be that the lower TDPs mean the chips are not boosting as high as Ryzen 7000 in games.

That's very unlikely, as long as ZEN5s power draw isn't massively up from ZEN4. 7700X needslike 70W in gaming, 9700X can draw 88W. Ryzen 9 SKUs are even further away from TDP limitations at gaming.

Looking at the core itself (as the platform is the same), there is no reason for Zen 5 to show mediocre uplift in games.

Problem is that the last years have taught us that Gaming isn't that much about the arch but more about Cache size and latency. I mean look at Rocket Lake. You can say what you want about it, but the IPC uplift definitely was there. In Gaming however, nothing. Why? L3 reduction from 20MB to 16MB and worse latencies.

AMD doubled effective L3 size in it's first 3 Gens.
ZEN: 4 Core CCX with 8MB L3.
ZEN2: 4 Core CCX with 16MB L3.
ZEN3: 8 Core CCX with 32MB L3.
All had sizable gaming uplifts. Then came ZEN4 that only changed L2. And ZEN at Spec RAM is ~20% faster than ZEN3 at Spec RAM. While having 15-16% higher clocks. You know what that means for gaming IPC, right? Now ZEN5 arrives with basically no changes to Cache and Clocks, so what do we think? I think an ~5% gaming uplift is definitely possible.

 

[Goop_reformed]

I would not bother to take what he said serious half of his posts are "Is zen5 good yet?" "So is it good yet?" "I've not recovered from the 32% hangover ." "So what's the consensus here, is zen 5 good yet?"

What's the deal with the invasion of freshly created accounts like this?

 

[leoneazzurro]

My source is a VAR channel which focuses on professional applications. Ram timing might play a role but I wasn't informed of that. Performance is exactly what amd has advertised minus a few percent.

RAM timings are quite important for gaming, especially when there is no 3D Cache. AMD results were obtained with good DDR5-6000 modules, IIRC. Also platform differences (I.e. different motherboards) may account for a few % in performance.

 

[Abwx]

so what do we think? I think an ~5% gaming uplift is definitely possible.

At Computerbase they have the 5800X3D 1% faster than the 7700X, so for you a 12% uplift over the 5800X3D would be only 5% above the 7700X..?..
What kind of math is this.?.

Intel Core i9-14900KS im Test: Benchmarks in Spielen vs. AMD Ryzen 9 7950X3D und Ryzen 7 7800X3D

Intel Core i9-14900KS im Test: Benchmarks in Spielen vs. AMD Ryzen 9 7950X3D und Ryzen 7 7800X3D / Testsystem und Testmethodik für Spiele

www.computerbase.de

 

[Philste]

At Computerbase they have the 5800X3D 1% faster than the 7700X, so for you a 12% uplift over the 5800X3D would be only 5% above the 7700X..?..
What kind of math is this.?.

CB tests with Spec RAM, and that is 5200MT for 7700X. AMDs test of 9700X was with 6000MT EXPO and most outlets that test like this (HUB for example) measured the 7700X to be around 8% faster than 5800X3D. That's why I said 4-5% with same RAM.

 

[coercitiv]

AMDs test of 9700X was with 6000MT EXPO and most outlets that test like this (HUB for example) measured the 7700X to be around 8% faster than 5800X3D. That's why I said 4-5% with same RAM.

AMD also tested with 7900XTX instead of 4090 used by most outlets. Keep in mind this can also skew results.

 

[Abwx]

CB tests with Spec RAM, and that is 5200MT for 7700X. AMDs test of 9700X was with 6000MT EXPO and most outlets that test like this (HUB for example) measured the 7700X to be around 8% faster than 5800X3D. That's why I said 4-5% with same RAM.

Biggest difference is that Computerbase is at 720p while HUB is at 1080p, beside the 7950XTX more or less give the same results as a 4090, so you are just doing some hasardous speculations based on things like this :

Problem is that the last years have taught us that Gaming isn't that much about the arch but more about Cache size and latency.

Have you first hand infos about Zen 5 cache latencies..?.

Because here you are saying that nothing was improved in this area, wich is just wild speculation to get to a desired point.

 

[Rheingold]

Also @SteinFG

That's been a core part of the Zen strategy from the beginning - reusing dies.

This re-usability was what saved then struggling AMD at the beginning of the Ryzen era. They designed just one die. And then they built Ryzen 1000 with it. Then they combined up to four of them to bring EPYC Naples. And when they realized there was demand for HEDT, they also combined two and later four of them to bring Threadripper. That's all the very same piece of silicon, called Zeppelin. What a marvel of cost effectiveness.

It's only with the switch to chiplets that EPYC got separate IOD variants.

 

[CouncilorIrissa]

The rumored changes to boost L3 density? Maybe latency got worse and DRAM tuning plays greater role

We've got AIDA64 GNR latency numbers, the latency is in line with Zen 4.

What's the deal with the invasion of freshly created accounts like this?

What's the deal with gatekeeping on a public forum?

 

[Nothingness]

Biggest difference is that Computerbase is at 720p while HUB is at 1080p, beside the 7950XTX more or less give the same results as a 4090, so you are just doing some hasardous speculations based on things like this :

Start of rant...

I don't understand the point of measuring CPU game performance at 720p or 1080p. This doesn't rule out driver deficiencies or GPU performance variations for resolutions no gamer uses. IMHO that's completely pointless and I'd prefer reviewers spend time analyzing results they get on purely CPU bound results rather than just listing useless figures with no or little analysis.

Also a much more interesting thing to test ASAP is RAM BW/latency sensitivity of a CPU so that early adopters pick the right RAM for their new platform.

And if they insist one measuring CPU/GPU performance they should use resolutions and a GPU that are likely to be paired with the tested CPU. Not easy, I know, but testing a high-end CPU with a high-end GPU on stupidly low resolutions is of no use to me.

 

[Abwx]

Start of rant...

I don't understand the point of measuring CPU game performance at 720p or 1080p. This doesn't rule out driver deficiencies or GPU performance variations for resolutions no gamer uses. IMHO that's completely pointless and I'd prefer reviewers spend time analyzing results they get on purely CPU bound results rather than just listing useless figures with no or little analysis.

The lower the resolution the more you ll be CPU bound, that s why Computerbase test at 720p, at 1080p/1440p CPUs differences s will be compressed.

Not easy, I know, but testing a high-end CPU with a high-end GPU on stupidly low resolutions is of no use to me.

Same as above.

 

[BorisTheBlade82]

Anyway I was wondering are decoders in Zen5 statically partitioned between SMT threads, so when SMT threads is enabled each thread ends up with 4 wide decoder and only if you disable SMT in the BIOS it might happen that both decoders will be trying to decode the same instruction stream?

According to the Mike Clark Interview with CnC, the first one is true. The Decoder indeed gets statically partitioned when two threads are allocated to one core.
The latter is not the case. According to him, one thread is able to make use of the full 2x4 decoders and there is no need to disable SMT in the BIOS. Of course, further constraints or limitations might apply.

 

[yuri69]

According to the Mike Clark Interview with CnC, the first one is true. The Decoder indeed gets statically partitioned when two threads are allocated to one core.
The latter is not the case. According to him, one thread is able to make use of the full 2x4 decoders and there is no need to disable SMT in the BIOS. Of course, further constraints or limitations might apply.

I'm really looking forward to any AMD docs detailing the constrains (non-align stuff, max length, etc.). The Family 1Ah SoG might be interesting.

 

[Nothingness]

The lower the resolution the more you ll be CPU bound, that s why Computerbase test at 720p, at 1080p/1440p CPUs differences s will be compressed.

Ha yes that's so obvious, I had never thought of that 😅 /s

You didn't address anything at all about my post.

As a gamer, I don't care how a new CPU performs at 720p/1080p on a 4090. I want to know how it will perform at resolution most gamers now play with a GPU that doesn't cost three times more than the CPU.

As a CPU designer and as a user of purely CPU bound software, I want to know how a new CPU performs on benchmarks that don't involve GPU or any other unrelated piece of hardware (NPU, etc.). I want deep analysis from a knowledgeable person, not some trash data dumped by a clown doing videos.

I wouldn't care about these 720p/1080p results with a $2k GPU if reviewers also spent time trying to understand their results. My feeling is that most of them just pile up results and are not able to analyze and understand what they get.

 

[BorisTheBlade82]

I'm really looking forward to any AMD docs detailing the constrains (non-align stuff, max length, etc.). The Family 1Ah SoG might be interesting.

Yepp, indeed. As I might not have the endurance to read through all that stuff, I am hoping for a nice summary by CnC, garnished with some nice micro-benchmarking 😉

 

[yottabit]

Ha yes that's so obvious, I had never thought of that 😅 /s

You didn't address anything at all about my post.

As a gamer, I don't care how a new CPU performs at 720p/1080p on a 4090. I want to know how it will perform at resolution most gamers now play with a GPU that doesn't cost three times more than the CPU.

As a CPU designer and as a user of purely CPU bound software, I want to know how a new CPU performs on benchmarks that don't involve GPU or any other unrelated piece of hardware (NPU, etc.). I want deep analysis from a knowledgeable person, not some trash data dumped by a clown doing videos.

I wouldn't care about these 720p/1080p results with a $2k GPU if reviewers also spent time trying to understand their results. My feeling is that most of them just pile up results and are not able to analyze and understand what they get.

You’re a CPU designer?

 

[Philste]

Interestingly enough, HUB just uploaded a Video basically comparing the RAM they (amd AMD) test with, with the one Computerbase tests with on an 7700X. With 6000 EXPO, 7700X is a whopping 12% faster. Doesn't look good for 9700X, right?

Oh and why can't I put Screenshots directly in my comments?