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

[DisEnchantment]

Speculate at will

 

[Philste]

Does that make any sense ? Think logically. I'm not saying it is impossible, but it is very unlikely the data represents retail products. Like others pointed out, there was a reason for the last minute recall.

It makes as much sense as any other data we have. Don't forget AMDs own benchmarks suite against 5800X3D, which clearly tells that 9700X is ~10% slower than 7800X3D. And that's 1st party benchmarks. So 15% slower in reality sounds reasonable. Apart from that: If all current samples are flawed, how can that random Chinese dude say they are 20% faster?

 

[inf64]

It makes as much sense as any other data we have. Don't forget AMDs own benchmarks suite against 5800X3D, which clearly tells that 9700X is ~10% slower than 7800X3D. And that's 1st party benchmarks. So 15% slower in reality sounds reasonable. Apart from that: If all current samples are flawed, how can that random Chinese dude say they are 20% faster?

Not true according to the same CB tests. 117/112= 1.044 or 4% slower than 7800X3D

 

[yottabit]

I guess another question about Strix Halo is if it will have the full AVX-512 implementation like desktop or the cut down Strix Point style AVX

 

[igor_kavinski]

I guess another question about Strix Halo is if it will have the full AVX-512 implementation like desktop or the cut down Strix Point style AVX

Most likely the latter. AVX-512 isn't used in any game AFAIK so no point in investing in those additional transistors for an APU.

 

[LightningZ71]

I wouldn't be totally surprised if the desktop APUs based on the strix/Kraken generation all use Kraken dies. The 16CU iGPU on Strix Point is even more memory bandwidth bound than Phoenix's 12CUs. Assuming that Kraken retains 8CUs and can clock them higher than Phoenix's 12CUs, the performance difference won't be much different using JEDEC standard, or even vanilla EXPO RAM.

That being said, we all see the growing market for the tiny form factor desktops that can't fit a GPU. Having an efficient APU with lots of cores and a high performance iGPU would certainly give more performance on the top end there. Assuming that they support memory overclocking, I can see the interest there. Without being heavily limited on thermals like a laptop, and having likely no real constraints on power delivery, it should be able to maintain higher clocks for both CPU cores and the iGPU as well. It should be close enough to low end desktop 3050 territory for performance while having more VRAM available.

 

[MS_AT]

Most likely the latter. AVX-512 isn't used in any game AFAIK so no point in investing in those additional transistors for an APU.

And is gaming the intended target or rather creative professionals that want some macbook with Windows? I mean I don't know and we haven't heard anything officially, but it would be a pity to let all this mem bandwidth go to waste

 

[Tuna-Fish]

Apart from that: If all current samples are flawed, how can that random Chinese dude say they are 20% faster?

They don't all need to be flawed for AMD to do a recall. If there was an issue that impacted half all manufactured chips, AMD would absolutely do a recall. They would do so for a lot lower percentage too. It's entirely possible that some of the random chips out there are fully functional.

The part that's sketchy to me is gaining that much from PBO. Recent AMD chips generally have not had a lot of OC headroom.

 

[Josh128]

Agree with Philste that the comments from the Chinese shop owner seem directly taken from AMD's tech day press slides. Only one that seems unique is the 9600X beating 9950X in some games. I wouldnt put much stock into anything he posted.

 

[Det0x]

Other day..

https://twitter.com/x/status/1816803713258025037

Not too shabby for a 6 core i would say 👍

 

[Josh128]

The part that's sketchy to me is gaining that much from PBO. Recent AMD chips generally have not had a lot of OC headroom.

Indeed. However, AMD recently admitted in an interview (think it was Mike Clark himself) that it was a challenge keeping power consumption of Zen 5 down on 4nm. Also, the rumor that 9950X draws 190W stock and doesnt reach 230W unless PBO is turned on seems very odd. If Zen 5 continues to scale perf very well up to 230W, why would that not be the default operation for this flagship SKU when it was for 7950X?

I could understand the "visuals" of making 9700X and 9600X 65W CPUs, but AMDs own tech day charts show that 65W TDP is obviously not enough to feed a 9700X to its full potential, because it gains more % from PBO than any other SKU. Very strange temp and power dynamics across the board for Zen 5 for based on the leaks we currently have.

 

[Josh128]

Other day..
View attachment 103906

https://twitter.com/x/status/1816803713258025037

Not too shabby for a 6 core i would say 👍

Oh wow. Thanks for posting. That single core score is considerably lower than what I would expect for the 5.4 boost clock. 7700X has same boost clock and gets ~2010. +17% IPC (claimed by AMD for R23) would put the score well into the 2300s. As it stands, vs CPUMonkeys 7600X 1976 ST score, this is only a 9.2% uplift.

All core uplift is also only 6% vs 7600X's 15.3K. Contrary to the "not too shabby" caption, this is objectively not a good look. Hmm...

 

[tsamolotoff]

f Zen 5 continues to scale perf very well up to 230W, why would that not be the default operation for this flagship SKU.

Lotsa people cried about 95C when Zen4 released, so that's maybe why. Also, maybe N4X(P) is not tuned for longevity with high temperatures, who knows.

 

[IronLynx]

put the score well into the 2300s.

He posted another! 2244 and 17037!

 

[yottabit]

And is gaming the intended target or rather creative professionals that want some macbook with Windows? I mean I don't know and we haven't heard anything officially, but it would be a pity to let all this mem bandwidth go to waste

Yeah I hope they had the foresight to go for that market. The M- chips have made great inroads in scientific computing community for “budget” options that either require CPU compute or large datasets that wont fit in average GPU VRAM

 

[Philste]

He posted another! 2244 and 17037!

That should be it for MT. 7600X has ~15300. AMD claims 17% uplift in Blender, where ZEN5 has 23% IPC. R23 has 17% IPC, so uplift should be around 12% -> 17100

 

[AcrosTinus]

With these performance numbers I might give AMD a try after the Phenom days. When are they going to move towards a active substrate ?

 

[gdansk]

With these performance numbers I might give AMD a try after the Phenom days. When are they going to move towards a active substrate ?

2026.

 

[Rheingold]

The part that's sketchy to me is gaining that much from PBO.

Dont' forget that Ryzen 9000 has lowered TDP for most of the models. All models below the 9950X basically default to what was the Eco mode before. Look at the right side of this slide:



The 9700X gains only 11% and not 22% because it's the SKU that's limited the most by its lower TDP compared to the 7000 series. Activating PBO does not help the 9950X very much, but the 9700X can gain the most.

 

[igor_kavinski]

The most interesting addition is the CheckMate benchmark, which emphasizes the capabilities of your CPU's branch prediction unit through a multi-threaded approach. The utility leverages the AVX (Advanced Vector Extensions) set, particularly the AVX, AVX2, and AVX-512, to accelerate certain types of calculations. More importantly, CheckMate has brought in support for AMD's newest Ryzen AI 300 APUs as well, which is certainly impressive since it will allow for more extensive testing of the new mobile platform.

New benchmark in latest AIDA64 and guess what we have from our best friend the ES user:



PBO CO daily tune.

 

[Abwx]

Activating PBO does not help the 9950X very much, but the 9700X can gain the most.

15% for the 9700X.

AMD unwraps Ryzen AI 300 series ‘Strix Point’ processors — 50 TOPS of AI performance, Zen 5c density cores come to Ryzen 9 for the first time

AI hits overdrive

www.anandtech.com

 

[Hail The Brain Slug]

15% for the 9700X.

AMD unwraps Ryzen AI 300 series ‘Strix Point’ processors — 50 TOPS of AI performance, Zen 5c density cores come to Ryzen 9 for the first time

AI hits overdrive

www.anandtech.com

Assuming this is just increasing TDP like it implies, that puts 9700X at ~28% faster than 7700X since the 7700X would see little to no gain just from increasing power.

 

[Rheingold]

15% for the 9700X.

Yep, that's the other slide. They pretty much complement each other.

 

[tsamolotoff]

New benchmark in latest AIDA64 and guess what we have from our best friend the ES user:

View attachment 103909

PBO CO daily tune.

 

[Abwx]

Assuming this is just increasing TDP like it implies, that puts 9700X at ~28% faster than 7700X since the 7700X would see little to no gain just from increasing power.

Possibly, because 7700X is the Zen 4 SKU that get the most in the inefficient part of the curve, even accounting 15W for the IOD the single CCD is still pushed at 115W, wich is more than the 7950X.

Yep, that's the other slide. They pretty much complement each other.

The results being "up to" this is also in Blender, but the number would still be at least 10% in CB and other such benches, they are really sure of themselves to reduce perf that much, they could had created an intermediary TDP between 65W/88W PPT and 105W/142W PPT for this SKU.

 

[StefanR5R]

Some interesting info from the Chinese shop owner who posted prices (sorry, link is to wccftech article)

"[...] With PBO, the Ryzen 7 9700X can beat the Ryzen 7 7800X3D [...]"

The part that's sketchy to me is gaining that much from PBO. Recent AMD chips generally have not had a lot of OC headroom.

Indeed. However, AMD recently admitted in an interview (think it was Mike Clark himself) that it was a challenge keeping power consumption of Zen 5 down on 4nm. [...]

This was not about absolute power consumption but in the context of co-developing Zen 5 for two different processes without letting the designs diverge too much from each other, to keep validation efforts in check.
(source, scroll down to the heading "The biggest challenge of Zen 5 design")

Anyway, the Chinese shop owner's claim ("...can beat...") is trivially true, as it merely depends on the game benchmark selection.