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

[DisEnchantment]

Speculate at will

 

[inf64]

Zen4 is like 30% slower than x3d in cs2, fyi, My own personal tests, done by locking the game to one or another CCD

Spoiler

x3d
29-09-2023, 21:20:29 cs2.exe benchmark completed, 56597 frames rendered in 72.796 s
Average framerate : 777.4 FPS
Minimum framerate : 647.1 FPS
Maximum framerate : 928.7 FPS
1% low framerate : 254.0 FPS
0.1% low framerate : 197.7 FPS
ccd1
29-09-2023, 21:22:08 cs2.exe benchmark completed, 42766 frames rendered in 69.109 s
Average framerate : 618.8 FPS
Minimum framerate : 324.9 FPS
Maximum framerate : 817.8 FPS
1% low framerate : 161.2 FPS
0.1% low framerate : 10.2 FPS
[/ISPOILER]

That is a massive difference that more cache brings. Thanks for sharing.

 

[soresu]

The total war games eat cpus for breakast.

Typically this means that the game engine is optimised like dog 💩

SWTOR was causing my CPU to ramp insanely when I first played it back in 2012 where just about every other game was doing fine.

 

[soresu]

That is a massive difference that more cache brings

It's pretty variable across different game engines and different levels tho from what I've seen of scores.

 

[leoneazzurro]

It's not ideal to compare 2CCD chip with buggy BIOS (as per his own admission) versus the best gaming chip on the market. It would have been great if he could have used at least 9700X in the comparison.

On a partial excuse, he said the BIOS is the latest available on the Gigabyte site for his motherboard, and it has the 9000 series support advertised. On the negative side, a recap page with the configuration and list of drivers used would have helped a lot.

 

[Josh128]

I mean I got about 8-9% lead for 7800X3D with these tests. That would match the "12% faster than 5800X3D" claim. So I would say yeah, seems to be legit. Now you can argue if the numbers are good or not. I would say they are close to horrible.

But what do the guys at WTFTech smoke to write "lower power" in the headline? 102W average is way above 7800X3D. However, it's the 12 Core so it's sameish to 7900X. Still looks like a gen to skip. If they can't get 3D to clock with normal clocks, they only will have like a 10% lead over ZEN4X3D. Probably not enough to convince the majority to upgrade.

Man, I need to go on a tech speculation guessing game or something, Ive been on fire lately, lol. I just got finished saying ~5% to 8% slower than Zen 4 X3D in gaming a few posts earlier, lol.

Hassan apparently only looked at the power consumption in one of the games, lol. That place has gone to hell in terms of article quality lately, even by its own low standards.

About the Zen 5 X3D, its almost a certainty that it will not hit normal clocks. I did predict that for Zen 4 X3D last go round and was totally wrong. This time, I would bet we'll probably see ~5.2 GHz out of the box, with AMD letting users play at their own risk trying to squeeze out another 100 to 200 MHz. I'd like to be pleasantly surprised, but unless something radically different was done in the design, I think thats what we'll get.

 

[Hail The Brain Slug]

On a partial excuse, he said the BIOS is the latest available on the Gigabyte site for his motherboard, and it has the 9000 series support advertised. On the negative side, a recap page with the configuration and list of drivers used would have helped a lot.

The "latest gigabyte bios" is buggy even with Zen 4 cpus. I have had to rollback from 1.2.0.0a to 1.1.7.0

 

[leoneazzurro]

Hmm.. it is not clear - the guy seems to be not an expert in memory config. He thinks that the fact Ryzen 9000 supports a IF clock of 2400MHz means that IF should run at that speed regardless of the memory speed. He also said he had to manually act on the IF speed because the scores for the 9900X were all over the place. Quite frankly, i think he speeded up too much for being the first reviewer but I'd look more at GN or HUB reviews for more appropriate memory settings.

 

[Philste]

Horrible? The 9900X is clearly not the best choice for the gaming, but this was clear even before the launch, just as the 7900X is not the best choice for the gaming.

Yeah, your right. I can't really hide that I expected more out of ZEN5. Not the 32% or whatever, but I still thought it would be a leap. I mean it's the first time AMD widened the arch like that since ZEN exists. And now it looks to be straightup worse than ZEN4 from the pure "performance uplift compared to predecessor" side. I mean Vanilla ZEN4 at least managed to be around 5800X3D. If 9700X is really 8-9% behind 7800X3D that is really a big gap, given we are at a point where 5% mean destroying the competition for some people.

I mean AMD showed the Blender generational uplift slides for Multithread and the best one was 9950X with 22%, worst one 9700X with 11%. And Blender was one of the biggest numbers in the IPC Slide, 23% if I remember right. So MT average is like 7-15% at average? ZEN4 had 30+% gaming also seems to be Single digit, at least compared to ZEN4 on same RAM.

 

[tsamolotoff]

IF clock of 2400MHz means that IF should run at that speed regardless of the memory speed.

If he manually tuned IF without testing for stability, he can get very random results. Also, sync mode is worth it only if frequency difference is very small (compared to desync), otherwise higher fclk will always beat or be on par with sync (7800 in 1:2 mode with fclk=1950 vs 2167 for example, and so on).

 

[Markfw]

Yeah, your right. I can't really hide that I expected more out of ZEN5. Not the 32% or whatever, but I still thought it would be a leap. I mean it's the first time AMD widened the arch like that since ZEN exists. And now it looks to be straightup worse than ZEN4 from the pure "performance uplift compared to predecessor" side. I mean Vanilla ZEN4 at least managed to be around 5800X3D. If 9700X is really 8-9% behind 7800X3D that is really a big gap, given we are at a point where 5% mean destroying the competition for some people.

I mean AMD showed the Blender generational uplift slides for Multithread and the best one was 9950X with 22%, worst one 9700X with 11%. And Blender was one of the biggest numbers in the IPC Slide, 23% if I remember right. So MT average is like 7-15% at average? ZEN4 had 30+% gaming also seems to be Single digit, at least compared to ZEN4 on same RAM.

NO retail CPUs, even to respected websites yet, and you are complaining about Zen 5 ? Why don't you wait until at least a respected site comes out with a review using retail chips ?

 

[Philste]

NO retail CPUs, even to respected websites yet,

How do we know? The point of NDAs is that they can't tell you about everything, that includes the date where they receive the CPUs. Review date should be in 7 Days, 8 days at worst. Reviewers should've already received their CPUs, everything else would be REALLY late.

 

[MS_AT]

Yeah, your right. I can't really hide that I expected more out of ZEN5. Not the 32% or whatever, but I still thought it would be a leap. I mean it's the first time AMD widened the arch like that since ZEN exists. And now it looks to be straightup worse than ZEN4 from the pure "performance uplift compared to predecessor" side. I mean Vanilla ZEN4 at least managed to be around 5800X3D. If 9700X is really 8-9% behind 7800X3D that is really a big gap, given we are at a point where 5% mean destroying the competition for some people.

I mean AMD showed the Blender generational uplift slides for Multithread and the best one was 9950X with 22%, worst one 9700X with 11%. And Blender was one of the biggest numbers in the IPC Slide, 23% if I remember right. So MT average is like 7-15% at average? ZEN4 had 30+% gaming also seems to be Single digit, at least compared to ZEN4 on same RAM.

Vanilla Zen4 had IPC boost, MHz boost and DRAM boost to overcome the 5800x3d. Zen5 has only IPC boost. As a bonus in this particular review it has misconfigured DRAM [and x3d will care less about misconfigured DRAM] and cross CCD penalty in somewhat well threaded games. While I don't expect 9900x will overcome 7800x3d I think the average scores for 9900x from other reviewers will be a tad higher due to those factors. The more interesting comparison from CPU vs CPU point of view would be 7900x vs 9900x

 

[Nothingness]

There s rarely reviews of CPU perfs with mainstream cards, you should complain to the reviewers.

I once did adress this point, one more time the fault is on the reviewers who never test realistic set ups, FI they test mainstream CPUs with the most expensive GPU, wich is a case that do not exist in real life, and likewise they test mainstream GPUs with the faster CPU, wich is also a case that barely exist in real set ups of consumers.

So I think we agree that many reviews are poor on that side.

So for the time we ll have to be content with AMD s 7900XTX based numbers wich are surely at 1080p, there s a computerbase review, at 720p, with numbers from this GPU in function of the CPU.

Ryzen 9 7950X3D im Test gegen 5800X3D und 13900K(S)

Zen 4 3D ist da. Im Test erweist sich AMDs Ryzen 9 7950X3D gegen 7950X, 5800X3D und Core i9-13900K(S) als schnell und effizient.

www.computerbase.de

I'll just keep on ignoring these parts of the reviews (and complain that reviewers should better spend their time on analyzing results of application benchmarks ).

 

[Det0x]

About the Zen 5 X3D, its almost a certainty that it will not hit normal clocks. I did predict that for Zen 4 X3D last go round and was totally wrong. This time, I would bet we'll probably see ~5.2 GHz out of the box, with AMD letting users play at their own risk trying to squeeze out another 100 to 200 MHz. I'd like to be pleasantly surprised, but unless something radically different was done in the design, I think thats what we'll get.

Above 5.35ghz fmax (stock) together with 2x L1, ~1.4x L2 and ~1.15x L3 cache bandwidth
Ontop of 7ns L3 and the core IPC gains going from Z4 to Z5

What do you think the final gaming performance uplift will be vs the 7800x3d ?

 

[csbin]

EPYC 9755 and 9965

https://twitter.com/x/status/1815764106781315276

https://twitter.com/x/status/1815764970191417456

 

[Elfear]

Above 5.35ghz fmax (stock) together with 2x L1, ~1.4x L2 and ~1.15x L3 cache bandwidth
Ontop of 7ns L3 and the core IPC gains going from Z4 to Z5

What do you think the final gaming performance uplift will be vs the 7800x3d ?

I'm not on fire with predictions like @Josh128 but I think we'll see 12-15% with Z5X3D vs Z4X3D.

 

[LightningZ71]

Yeah it's called Turin Dense and it's on N3e.

What's that

I thought that Turin Dense was using Zen5C cores, not Zen5... not that there's a functional logic difference between them, so that's a only a technicality?

 

[Rheingold]

The retail CPUs and reviews are not even out yet

Well, at least the linked video *is* a review of a retail CPU. Once the SKUs reach the dealers, then in the days before launch some samples always find their way to outlets that didn't sign an NDA.

 

[Gideon]

That will extend existing DDR5 infrastructure for a while longer, but CAMM is just plain superior.
DDR6 has a big baseline cost associated and really is tailored for servers and nothing else.

it's just 1/3 of registered memory (buffering clocks, but not data) and it's also coming to the CAMM2 modules. From the article:

In any case, CKDs will be coming to all of JEDEC's DDR5 memory form factors. So along with the CUDIMM, we'll have the Clocked SODIMM (CSODIMM), and even DDR5 CAMM2 memory modules will use clock drivers.

 

[Nothingness]

You’re a CPU designer?

Yes, kind of. I'm writing accurate CPU simulators to study ideas to improve the performance of upcoming designs. I'm now more involved in the infrastructure of the simulators and in studying the benefit of new instructions rather than, for instance, studying details of branch prediction or data prefetchers. That's why I'm much more interested in the deep study of results of CPU bound benchmarks rather than the performance of the CPU/GPU couple at 720p.

And no, as @igor_kavinski wrote, I won't tell what companies I work or worked for 😀

 

[Joe NYC]

I'm not on fire with predictions like @Josh128 but I think we'll see 12-15% with Z5X3D vs Z4X3D.

That would be proportional to 16% IPC uplift, and assume that Zen 5 x3d has the same clock speed deficit vs non x3d as Zen 4.

 

[lightmanek]

Yes, kind of. I'm writing accurate CPU simulators to study ideas to improve the performance of upcoming designs. I'm now more involved in the infrastructure of the simulators and in studying the benefit of new instructions rather than, for instance, studying details of branch prediction or data prefetchers. That's why I'm much more interested in the deep study of results of CPU bound benchmarks rather than the performance of the CPU/GPU couple at 720p.

And no, as @igor_kavinski wrote, I won't tell what companies I work or worked for 😀

I could have a pretty decent guess at which comanies they might be as my good friend most likely works for one of them too doing exactly what you described.

Cool stuff and the product is amazing 👏

Back on topic, Zen5X3D 20% faster in games than Zen4X3D is my shot in the dark.

 

[Hail The Brain Slug]

I could have a pretty decent guess at which comanies they might be as my good friend most likely works for one of them too doing exactly what you described.

Cool stuff and the product is amazing 👏

Back on topic, Zen5X3D 20% faster in games than Zen4X3D is my shot in the dark.

If Zen5X3D keeps the same clocks as Zen 5, it should be quite a lot faster. Zen3X3D and Zen4X3D both had a pretty big handicap in clockspeeds.

It seems like everyone's focusing on the clockspeed stagnation on vanilla Zen 5 parts being the reason for lesser game uplift. But the X3D parts will likely see a big clockspeed uplift. Giving them a pretty good advantage.

 

[Markfw]

Well, at least the linked video *is* a review of a retail CPU. Once the SKUs reach the dealers, then in the days before launch some samples always find their way to outlets that didn't sign an NDA.

I don't believe its a retail CPU, or other respected websites would have reviews too. Could be an ES. You can relink for me if you want, I did not see it. If its a youtube, then its not worth even reading.

 

[CakeMonster]

Kind of a disappointing leak, getting your hands on a new gen chip, and testing with weird non-standard memory and IF settings so that no one comes away wiser.