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

[DisEnchantment]

Speculate at will

 

[Ghostsonplanets]

Yea.

4+4. Chop is 2+4.

Yeah 8 CUs.
And a full AIE because le AIPC maymay.

RPL 282 is still a far cheaper option but not compelling in 2025 at all.

These are actually really great. 4 + 4 and 8CU will be enough for mainstream consumers for the foreseeable future.

 

[CouncilorIrissa]

These are actually really great. 4 + 4 and 8CU will be enough for mainstream consumers for the foreseeable future.

Will be interesting to see how the 8CU RDNA3.5 stacks up against the 780M in handhelds.

 

[adroc_thurston]

Will be interesting to see how the 8CU RDNA3.5 stacks up against the 780M in handhelds.

Pretty well, PHX iGP is a mess.png

 

[CouncilorIrissa]

Pretty well, PHX iGP is a mess.png

You mean it's a mess in the same way every other RDNA3 product is a mess, or is there anything else wrong with it?

 

[adroc_thurston]

You mean it's a mess in the same way every other RDNA3 product is a mess

Yes.
Pretty obvious when you lock PHX1 to 12-15W and load any remotely CPU+GPU-intensive game.
The thing just falls apart in really funny ways.

I think AC Valhalla showcases it in a particularly pronounced way.

 

[CouncilorIrissa]

Yeah, it's lower power behaviour is sort of dodgy, but I wasn't sure whether it's something unique to PHX1 or common for anything-not-Van Gogh.

But I think I saw some graphs on X suggesting that HWK1 is a bit better in this regard.

 

[adroc_thurston]

But I think I saw some graphs suggesting that HWK1 is a bit better in this regard

A bit but not enough.

 

[CouncilorIrissa]

A bit but not enough.

Would be nice to have some alternatives, but as it stands they're competing against themselves with MTL-U not looking too hot (or somewhat warm really).

 

[PJVol]

Ehhh, making MS teams even remotely snappy is the end goal.

I have a little idea how big this segment is, but whatever it is I assume you meant JS client libs and not the core SDKs backend.

 

[adroc_thurston]

but as it stands they're competing against themselves with MTL-U not looking too hot (or somewhat warm really).

LNL looks good for ultra low power stuff.
But $$$.

 

[BTRY B 529th FA BN]

Isn't Fortnite CPU heavy over GPU? A lot of the UT4 players I play with all say it's CPU

 

[adroc_thurston]

Isn't Fortnite CPU heavy over GPU?

Not at Epic settings.
Maxed out Lumen and Nanite nuke all GPUs available on the market for orbit.

 

[Goop_reformed]

No, the poster was implying the average IPC is 3x% ( meaning between 30 and 39%). Clocks seem to be roughly the same or a slightly higher for ST with Zen 5. Overall it looks like a total pwnage.

Monthly corrections + plus "I told you so" incoming. There's also a recent video about the number 37 so I'll take 37% ipc uplift

However, heat is a major for me and for many others as the 7800x3d is just that good/cool.

 

[adroc_thurston]

However, heat is a major for me and for many others as the 7800x3d is just that good/cool.

Just toggle the ecomode.
It's a literal one press PPT down button.
Can't be any simpler.

 

[Joe NYC]

Isn't Fortnite CPU heavy over GPU? A lot of the UT4 players I play with all say it's CPU

I came across this. Looks like it is nearly impossible to move the score by more than 1-2% due to GPU limitation, and Zen 5 goes up by 7%

 

[gdansk]

Why wouldn't they use CPU bound settings for testing CPU performance?

 

[H433x0n]

Interesting that GB6 wasn’t shared by AMD. That’s the most representative benchmark (IMO) and would give a good idea of the overall performance.

 

[Tigerick]

IPC Uplifts are not the same as Performance Numbers

	CB R23 1T	Passmark	7-Zip	V-Ray	Dolphin	wPrime (1024M)
Zen 3 5950X (3.4 - 4.9GHz, 105W)	1655	3468	165748	29738	227	49.434
Zen 4 7950X (4.5 - 5.7 GHz, 170W)	2066.8	4290	209616	42367	170	30.957
+ %	+ 25%	+ 24%	+ 26%	+ 42%	+ 33.5%	+ 60%
Zen 4 IPC @ 4GHz Average +13%	+ 9%	+ 5%	+ 13%	+ 15%	+ 32%	+ 39%
Zen 5 16-core (4.3 - 5.7 GHz, 170W)
Zen 5 IPC @ ? GHz Average +16%	+ 17%

I can't believe people don't understand the difference between performance numbers and IPC uplifts. As shown in the table above, performance numbers of Zen4 are far better than Zen3 and yet the IPC uplifts tested by AMD shown average 13% better uplifts. The reason is because Zen4 is clocked much higher than Zen3 thus consume more power.

IPC of Zen5 would be much lower, my estimate IPC would be around 15-20% faster...Definitely not 40% uplift as adroc claimed.

@adroc_thurston Do you still think your 40% is IPC or performance numbers?

 

[adroc_thurston]

Interesting that GB6 wasn’t shared by AMD

Oh but it is, in Strix disclosure no less.

That’s the most representative benchmark (IMO)

That would be some JavaScript hellware.
GB6 mixes a whole pile of completely unrelated workloads (2/3rds aren't even client at all).

 

[gdansk]

Interesting that GB6 wasn’t shared by AMD. That’s the most representative benchmark (IMO) and would give a good idea of the overall performance.

For Zen 4? AMD haven't shared anything for Zen 5 yet. But GB6 wasn't out yet when Zen 4 launched.
GB6 shows Zen 4 as 35%-40% better 1T than Zen 3 while generally it's closer to 25%. Normalize for clock rate and it's still an outlier making Zen 4 look better than it is.

 

[Tigerick]

It s just that CB starting from R20 is the best case for Intel among all tests along with Povray wich use AVX2 only for Intel, let s look at 7 ZIP wich is integer code and more representative of consumer apps, you ll be surprised.

There is leaks about 7-Zip numbers from Allthewatts, 28% faster score is 8 percent faster. I am not surprised cause I believe IPC numbers would be lower. What do you think?

 

[H433x0n]

Somebody pointed out to me that it’s probably fake.

All of those numbers are the Zen 4 figures but multiplied by 2.2x (which is conveniently 30%).

 

[Goop_reformed]

Somebody pointed out to me that it’s probably fake.

All of those numbers are the Zen 4 figures but multiplied by 2.2x (which is conveniently 30%).

Did that to avoid getting bonked

 

[Kaffeekenan]

IPC Uplifts are not the same as Performance Numbers
View attachment 96183

	CB R23 1T	Passmark	7-Zip	V-Ray	wPrime (1024M)
Zen 3 5950X (4.9GHz, 105W)	1655	3468	165748	29738
Zen 4 7950X (5.7 GHz, 170W)	2066.8	4290	209616	42367
+ %	+ 25%	+ 24%	+ 26%	+ 42%
Zen 4 IPC 4GHz Average +13%	+ 9%	+ 5%	+ 13%	+ 15%	+ 39%
Zen 5 16-core (? GHz)	+ 20%	+ 11%	+ 28%	+ 33%	+ 86%
Zen 5 IPC 4GHz Average ?

I can't believe people don't understand the difference between performance numbers and IPC uplifts. As shown in the table above, performance numbers of Zen4 are far better than Zen3 and yet the IPC uplifts tested by AMD shown average 13% better uplifts. The reason is the because Zen4 is clocked much higher than Zen3 thus consume more power.

The same will happen to Zen5; I am sure AMD will try to clock Zen5 higher, hence the leaked results are ranged from 11% to 86%. But IPC of Zen5 would be much lower, my estimate IPC would be around 15% faster...Definitely not 40% uplift as adroc claimed.

@adroc_thurston Do you still think your 40% is IPC or performance numbers?

Zen 5 and Zen 4 will be clocked almost the same so you will not see the big clockboost like from Zen 3 to Zen 4. So IPC and performance will be kinda close with Zen 5. Cause more performance is caused by more ipc and more clocks...

 

[CouncilorIrissa]

IPC Uplifts are not the same as Performance Numbers
View attachment 96183

	CB R23 1T	Passmark	7-Zip	V-Ray	wPrime (1024M)
Zen 3 5950X (4.9GHz, 105W)	1655	3468	165748	29738
Zen 4 7950X (5.7 GHz, 170W)	2066.8	4290	209616	42367
+ %	+ 25%	+ 24%	+ 26%	+ 42%
Zen 4 IPC 4GHz Average +13%	+ 9%	+ 5%	+ 13%	+ 15%	+ 39%
Zen 5 16-core (? GHz)	+ 20%	+ 11%	+ 28%	+ 33%	+ 86%
Zen 5 IPC 4GHz Average ?

I can't believe people don't understand the difference between performance numbers and IPC uplifts. As shown in the table above, performance numbers of Zen4 are far better than Zen3 and yet the IPC uplifts tested by AMD shown average 13% better uplifts. The reason is the because Zen4 is clocked much higher than Zen3 thus consume more power.

The same will happen to Zen5; I am sure AMD will try to clock Zen5 higher, hence the leaked results are ranged from 11% to 86%. But IPC of Zen5 would be much lower, my estimate IPC would be around 15% faster...Definitely not 40% uplift as adroc claimed.

@adroc_thurston Do you still think your 40% is IPC or performance numbers?

In order to get 40% perf uplift with 15% IPC gains you'd need a whopping 21% clock speed increase. That would mean a literal 5.7*1.21 = 6.9 GHz (lol) for the top SKU, on the same node no less.