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

[DisEnchantment]

Speculate at will

 

[Hitman928]

Shouldn’t that affect the fp score as well? Seems particular to just int.

~5, percent?

I'm making some assumptions that may turn out to be wrong, but if they are correct assumptions, then that would just mean that FP increase is higher than Anandtech shows.

 

[Abwx]

There's a problem with that Math and that Problem is ZEN5c. Clocks of 8 of the 12 Cores are Limited to 3.7GHz. Computerbase got an avg core clock of 3.43GHz at 33W. So starting at 40W, there's only 4 of 12 Cores that can scale further. So basically at that point every scaling is gone. Even if those 4 Cores got to 5.0 GHz that's only a 11-12% uplift over 40W. No matter how much power you throw in there. 1400 isn't possible.

GPD displayed a run that scored 1525, so that s outright possible, beside the 8500G APU wich is 2 + 4, so half the 370, can go above 45W according to Phoronix review.

AMD's Radeon 890M RDNA 3.5 GPU Obliterates 780M In Early Benchmarks

GPD, of GPD Win fame, has just dropped some early benchmarks for AMD's Strix Point, and they look stellar.

hothardware.com

 

[Rheingold]

I also added a fitted line for the HX 370 and X Elite, though it's just eyeballed and the lower power levels for the X Elite numbers are weird. I'm guessing it's different SoCs at the bottom, similar to the M3 issue.

The darker purple HX 370 ones are from two different models, one more geared to being quiet, one towards higher performance. The X Elite numbers are estimations since HWiNFO still has no way of reading package power for the Qualcomm chips. So JustJosh had to go from wall power readings and then substract estimates for the notebook baseline based on models as similar as they could find. But they say that the yellow values are probably best case values for the X Elite.

The main conclusion of the video is "If you're not buying Apple, get Zen5". Strix clearly makes the X Elite obsolete. CPU efficiency is just as good or better with Strix, and the performance advantage (and compatibility) of the iGPU is even bigger. And at this point we don't need to talk any more about Intel until Lunar Lake arrives.

 

[Hitman928]

I have doubts for actually ran this themselves they just copied those numbers from cpu-monkey who just applied amd 17% ipc uplift to zen 4 numbers

Good catch, those are just CPU Monkey numbers which are known to be made up, specifically for pre-release hardware. We have actual tests now that show that 1525 is not happening. Even 1400 is out of reach.

 

[Abwx]

I have doubts for actually ran this themselves they just copied those numbers from cpu-monkey who just applied amd 17% ipc uplift to zen 4 numbers

That s the other way around, CPU Monkey collect numbers here and there that they use for their database, not the first time they did so, they often use numbers that are allegedly leaked even before there s proved tests.

And in case you dont know GPD manufacture devices using AMD APUs,
why would they need to rely on obscure sites since they have ESs long before the consumers have access to actual products..?.

 

[ErichZannMusician]

So it is indeed Zen5%.

232 mm2 for this crap - barely any resources increase for the iGPU, NPU hobbling die space that could have been allocated to another 4 Zen5 cores, and a weird floorplan which unless there is an outer ring bus, would obliterate latency between the CCXs.

Meh.

 

[Josh128]

Good catch, those are just CPU Monkey numbers which are known to be made up, specifically for pre-release hardware. We have actual tests now that show that 1525 is not happening. Even 1400 is out of reach.

View attachment 104061

Guys you may have it backwards--I believe CPU Monkey used that leak and added to its site, not 100% but they did the same thing after a few other leaks.

 

[inf64]

Some very odd results, especially Specint in AT review. Zen 5 does look underwhelming considering AMD's claims.

 

[Rheingold]

would obliterate latency between the CCXs.

Sorry that I have to repeat-feature that in this thread, but David Huang already measured bad latencies:



That's worse than with Ryzen 1000 (had to search a bit, the earlier discussion of this was from PC Perspective):

 

[igor_kavinski]

Strix clearly makes the X Elite obsolete.

Cheaper (50% price drop!) Elite X laptops for us to mess with. Win-Win!

 

[igor_kavinski]

Zen 5 does look underwhelming considering AMD's claims.

It's not real Zen 5. It was castrated slightly to make space for the NPU.

 

[Hail The Brain Slug]

Guys you may have it backwards--I believe CPU Monkey used that leak and added to its site, not 100% but they did the same thing after a few other leaks.

I am inclined to agree this is probably what happened. CPU Monkey will aggregate results they find anywhere and everywhere - It seems much less likely an actual OEM system manufacturer would just peel numbers off CPU monkey.

 

[Philste]

GPD displayed a run that scored 1525, so that s outright possible, beside the 8500G APU wich is 2 + 4, so half the 370, can go above 45W according to Phoronix review

That's Marketing Material. With the 3rd Party reviews from today, 1500 should be impossible. We will see. But 1500 would be literally around 5.0GHz on all 12 Cores, which isn't possible. And I don't say the CPU can't go above 40W, I said scaling will be dead above that value, because only 4 of 12 cores can increase their clock. So for every % perfomance, the Cores need to increase the clock by 3 times as much as before.

 

[Hitman928]

Guys you may have it backwards--I believe CPU Monkey used that leak and added to its site, not 100% but they did the same thing after a few other leaks.

Either way, release hardware reviews trump any leaked results.

 

[Philste]

It's not real Zen 5. It was castrated slightly to make space for the NPU.

I doubt that FPU will make such a big difference in general Purpose. It's still odd for me that ZEN5 doesn't equal ZEN5 depending on the product.

 

[ErichZannMusician]

Nearly 2 GHz fMax difference between Zen5 and Zen5c cores.

Core-to-core latency within the 4 core Zen5 CCX worse than that of the 8 core CCX in Phoenix and Rembrandt.

Meteor Lake latency all the way to the LP-E cores is better than this.

 

[igor_kavinski]

It's still odd for me that ZEN5 doesn't equal ZEN5 depending on the product.

I think I read that Mike Clark said they worked on four different Zen 5 designs?

 

[SarahKerrigan]

I think I read that Mike Clark said they worked on four different Zen 5 designs?

z5c-256, z5c-512, z5-256, z5-512

 

[Abwx]

That's Marketing Material. With the 3rd Party reviews from today, 1500 should be impossible. We will see. But 1500 would be literally around 5.0GHz on all 12 Cores, which isn't possible. And I don't say the CPU can't go above 40W, I said scaling will be dead above that value, because only 4 of 12 cores can increase their clock. So for every % perfomance, the Cores need to increase the clock by 3 times as much as before.

IIRC AMD said that they extended the frequency range of the C cores, indeed they are only 25% smaller than the regular cores where Zen 4C was 35% smaller than its counterpart.

 

[Hitman928]

Nearly 2 GHz fMax difference between Zen5 and Zen5c cores.

Core-to-core latency within the 4 core Zen5 CCX worse than that of the 8 core CCX in Phoenix and Rembrandt.

Meteor Lake latency all the way to the LP-E cores is better than this.

Probably a trade off for efficiency.

 

[ryanjagtap]

I think I read that Mike Clark said they worked on four different Zen 5 designs?

I think so....
STX with both Zen 5 and Zen 5c with 2x256 AVX 512.
GNR Zen 5 with full width AVX 512
Turin Dense with Zen 5c with full width AVX 512
STX Halo with Zen 5 cores and whatever variation of Zen 5 they put in the SoC die

 

[naukkis]

should amd have just designed a 12 core ccx

Or just stay with 4+4 or even 6+2 configuration with same CCX. Now they have "12" core cpu where most of cores are pretty useless.

 

[GTracing]

Or just stay with 4+4 or even 6+2 configuration with same CCX. Now they have "12" core cpu where most of cores are pretty useless.

How are they useless?

 

[Philste]

IIRC AMD said that they extended the frequency range of the C cores,

David Huang leaked 3.3GHz clocks for 365 and 3.7GHz for 370 weaks ago. Yesterday @StinkyPinky confirmed the 3.3GHz for 365 with his retail unit, so there is no reason to believe the 3.7GHz for 370 are wrong. I wish they were tho.

 

[ryanjagtap]

should amd have just designed a 12 core ccx

Yup. Instead they made a modular IF which they can turn parts off to save space energy.