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

[DisEnchantment]

Speculate at will

 

[Hitman928]

it's not 2GHz.
what are you on.
clk readouts are in .gb5. like, wat?

.gb5 is reporting 2 GHz in the score we are discussing, so what are you talking about?

 

[Ghostsonplanets]

.gb5 is reporting 2 GHz in the score we are discussing, so what are you talking about?

He's talking about the 1.4GHz with ~1400 ST score benchmark

 

[adroc_thurston]

.gb5 is reporting 2 GHz in the score we are discussing, so what are you talking about?

Only the last, the recently posted one, is relevant.

 

[Henry swagger]

Zen 4 has slightly higher IPC than Golden Cove so the numbers track with the above calculation, roughly speaking at least. There’s basically no way Zen 5 gets that low of an increase unless AMD completely botched something. The far more plausible answer is that it’s just that the score is an ES on an immature platform that is limiting performance.

Zen 5 is to wide to have lower ipc increase tham zen 4 over zen 3

 

[adroc_thurston]

Zen 5 is to wide to have lower ipc increase tham zen 4 over zen 3

Wider cores don't automatically bump IPC by that much, see Cortex-X4.

 

[Henry swagger]

Especially when all prior Zens were 6 wide front ends and it's now 8 wide...if they bother to change something that big, I really don't get how they'd get a tiny tick level increase.

Or maybe the meme was just a way to cope and it's really Zen 5%.

Wait for 0ther benchmarks.. geekbench can be easily tricked as chips and cheese showed

 

[Hitman928]

He's talking about the 1.4GHz with ~1400 ST score benchmark

I know, that’s not the one @tamz_msc was using though and hence the discussion, so the comment about it not running at 2 GHz is wrong.

 

[tamz_msc]

Only the last, the recently posted one, is relevant.

Explain why.

 

[adroc_thurston]

Explain why.

Earlier samples have a pile of chicken bits popped off for post-Si validation purposes.

 

[tamz_msc]

Earlier samples have a pile of chicken bits popped off for post-Si validation purposes.

Yeah, right... The first one is in the morning, and the second one is in the afternoon.

 

[adroc_thurston]

Yeah, right... The first one is in the morning, and the second one is in the afternoon.

Coping already? Wow, fast.
it's ok bro Intel will be competitive with z5 in 2026. hopefully!

 

[tamz_msc]

Coping already? Wow, fast.
it's ok bro Intel will be competitive with z5 in 2026. hopefully!

I can literally see the timestamp on the entries.

It's about time to pull the curtain on your make-believe.

 

[Hitman928]

Only the last, the recently posted one, is relevant.

If the 1.4 GHz score is legit, I get roughly 70% IPC increase in integer. Seems too good to be true but we’ll see soon enough I guess.

 

[adroc_thurston]

I can literally see the timestamp on the entries.

It's about time to pull the curtain on your make-believe.

?
It's okay, 2026 will be your year!

If the 1.4 GHz score is legit, I get roughly 70% IPC increase in integer. Seems too good to be true but we’ll see soon enough I guess.

Remember that IPC/freq scaling isn't quite linear.
So at prod clocks it'll be lower.

 

[Hitman928]

?
It's okay, 2026 will be your year!

Remember that IPC/freq scaling isn't quite linear.
So at prod clocks it'll be lower.

Sure, this is all just rough math, but even in ballpark terms it seems too high. If true though, it’s hard to see Intel catching up, even by 2026.

 

[adroc_thurston]

Sure, this is all just rough math, but even in ballpark terms it seems too high

Well yeah, it's an impressive (and a bit crackhead) core.

If true though, it’s hard to see Intel catching up, even by 2026.

NVL should be promising enough.

 

[gdansk]

cpupower set-frequency doesn't seem to work on Zen 4. So I can't get 2000MHz results.

And while it's fun I'm pretty sure it's pointless to compare to possibly accurate GB5 results at nowhere near retail clock speeds.

 

[tamz_msc]

It's okay, 2026 will be your year!

The *timestamps* are there on the entries

 

[adroc_thurston]

The *timestamps* are there on the entries

No worries, 2026 will be your year!
hopefully.

And while it's fun I'm pretty sure it's pointless to compare to possibly accurate GB5 results at nowhere near retail clock speeds.

Yeah.
Useful enough for ballpark perf modeling though.

 

[tamz_msc]

cpupower set-frequency doesn't seem to work on Zen 4. So I can't get 2000MHz results.

And while it's fun I'm pretty sure it's pointless to compare to possibly accurate GB5 results at nowhere near retail clock speeds.

Does bash not ask you to install it?

 

[gdansk]

Does bash not ask you to install it?

It runs but all it does is disable turbo. It doesn't limit the cores frequency properly.

 

[tamz_msc]

It runs but all it does is disable turbo. It doesn't limit the cores frequency properly.

You have to set both an upper and lower frequency with -u and -d.

 

[gdansk]

You have to set both an upper and lower frequency with -u and -d.

I did the same steps as I did for the 12600K. 7800X3D doesn't seem to listen. I give up for now.

 

[CouncilorIrissa]

Benchmarking on Linux comes with all sorts of interesting caveats tbh. Performance may vary wildly depending on the cpu governor, scheduler, kernel timer frequency, you name it.

I was able to achieve swings ranging from -50% up to +15% of 1T performance in GB5 by tweaking these things on my old Whiskey Lake laptop compared to stock settings.

 

[adroc_thurston]

Benchmarking on Linux comes with all sorts of interesting caveats tbh. Performance may vary wildly depending on the cpu governor, scheduler, kernel timer frequency, you name it.

I was able to achieve swings ranging from -50% up to +15% of 1T performance in GB5 by tweaking these things on my old Whiskey Lake laptop compared to stock settings.

yeah Linux benchmarking is a mess outside of shelf stable server platforms.