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

[DisEnchantment]

Speculate at will

 

[StefanR5R]

Disable V-cache and re-run, if you can.

impossible, you can't disable v-cache

It can be done on Genoa according to AMD's "BIOS & Workload Tuning Guide for AMD EPYC™ 9004 Processors". Not on Raphael? Maybe it's just that your BIOS vendor does not expose this option.

However, the extra cache doesn't matter in image rendering benchmarks anyway. (source)

 

[Josh128]

Yeah but:

You don't even need to tuned it: According to launch reviews a 7950X just set to Eco Mode (142W PPT) loses only 5% performance compared to 230W PPT. So if you manually tuned it even further it's probably better than that ZEN5 sample.

Because the correct memory Channel Reporting came with 6.3. Your first screenshot is 6.2.1

Understand, but tuning it makes it even faster--ie the 9950X is specifically cited to be CO tuned + 120W PPT. So it has every possible advantage, but still seems to be no better than the gent above's 7950X @ 120W PPT?

 

[Josh128]

Probably it has close to zero impact, as mem bw is fairly low during the rendering itself (idk, ~5gbps ish from hwinfo metrics)

I was thinking you were talking about Geekbench runs, not Blender.

 

[igor_kavinski]

The ES clocks are quite a bit lower.

 

[MarkPost]

View attachment 102742

View attachment 102743

Still faster, it seems. Or i'm doing something wrong. Typical in-benchmark clocks on the first screenshot

I guess you are not running 3.3 version?

This is my 7950X3D (Blender 3.3 bench, SMT OFF, PPT 120W, CO tuned)

It seems to me, Zen 5 AVX512 full implementation has a great impact here (I guess Blender uses AVX512)

 

[Hail The Brain Slug]

I guess you are not running 3.3 version?

This is my 7950X3D (Blender 3.3 bench, SMT OFF, PPT 120W, CO tuned)

It seems to me, Zen 5 AVX512 full implementation has a great impact here (I guess Blender uses AVX512)

I thought we already concluded that SMT was not off, it was just a special mode of hwinfo?

 

[Josh128]

Theres zero chance that 9950X run was with SMT off.

 

[Det0x]

I guess you are not running 3.3 version?

This is my 7950X3D (Blender 3.3 bench, SMT OFF, PPT 120W, CO tuned)

It seems to me, Zen 5 AVX512 full implementation has a great impact here (I guess Blender uses AVX512)

Can you post scores with SMT enabled to compare against the other 7950X3D ? 👍

 

[MarkPost]

I thought we already concluded that SMT was not off, it was just a special mode of hwinfo?

IMO ES sample scores fits with SMT OFF, not a hwinfo special mode.

 

[MarkPost]

Can you post scores with SMT enabled to compare against the other 7950X3D ? 👍

here it is:

 

[Det0x]

here it is:

Thanks for confirming, both numbers for the 7950X3D seems about right then 👍 (+- ~4%)
If this continues, it will be interesting to see what happens at the big-end, although i suspect the X3D is drawing close to its limit
Would also be nice to see a regular 7950X, maybe brain slug will provide when he get the time

 

[Hail The Brain Slug]

Thanks for confirming, both numbers for the 7950X3D seems about right then 👍 (+- ~4%)
If this continues, it will be interesting to see what happens at the big-end, although i suspect the X3D is drawing close to its limit
Would also be nice to see a regular 7950X, maybe brain slug will provide when he get the time

Doing it on the side right now, have to install w11 again

 

[MarkPost]

Theres zero chance that 9950X run was with SMT off.

really? do you think Zen 5 is slower than Zen 4?

We come back to -5% IPC

 

[Markfw]

when whoever who gets the first benchmark out, please create a separate thread for it !

 

[Det0x]

really? do you think Zen 5 is slower than Zen 4?

We come back to -5% IPC

We are not really comparing "IPC" here, we are looking at power/performance atm..
And do keep in mind tsamolotoff which is a fellow hwbot user, is running among the fastest tuned 7950x3d's in the world, that is if i remember correctly and dont mix up ppl (?)
(lets just say 102.5 ECLK 7950X3D is not for everyone)

So far it seems like the Zen4X3D and Zen5 ES results are about even at the same PPT targets, which i would say is good thing (although the Z5 beat your Z4X3D at 120w example above)
But do note the clockspeed difference.. We already know what happens when Zen4 get above 5000mhz allcore, will same thing happen at 4000mhz for Zen5 or will it keep going ? 🤔

tsamolotoff's 7950X3D clocks while running the benchmark




PS: to the guys taking screenshots with 7950X3D and 7950X, if you reset hwinfo 1sec after you have started blender and take screenshot the very second benchmark ends, you wont pollute your hwinfo data.
(hwinfo updates with idle numbers like 1-2 seconds after blender benchmark ends)

PS2: we're also waiting for vanilla Zen4 numbers as its not clear for me if v-cache sees any benefit in blender.

 

[gaav87]

really? do you think Zen 5 is slower than Zen 4?

We come back to -5% IPC

Your clocks are 12% higher then that 9950x ur score is 1.5% higher. This puts us at 10.5%+ for zen5 still, a farcry from 23% on amd slides.

 

[tsamolotoff]

bclk overclock ?

A bit of eclk compensated with -fmax, my cpu can't do above 5750 in ST.

It can be done on Genoa according to AMD's "BIOS & Workload Tuning Guide for AMD EPYC™ 9004 Processors". Not on Raphael? Maybe it's just that your BIOS vendor does not expose this option.

There is no such option in the bios (and pretty sure Genoa-x3d is different as there can be more than 1-hi stacks of cache there)

IMO ES sample scores fits with SMT OFF, not a hwinfo special mode.

Pretty doubtful, as it'd mean zen5 is 50% faster iso-clock (although who knows, it seems igor's friend is a bit toying with us )

And do keep in mind tsamolotoff which is a fellow hwbot user, is running among the highest tuned 7950x3d's in the world, that is if i remember correctly (?)

No, your results are typically higher, I only scored 2-4th and the list of participants is quite short to be fair. My CPU can't score much above 38.5k in CB23, although maybe with finer CO tuning and clean OS instead of 'older than mammoth s***' system it could do more

 

[Abwx]

So far it seems like the Zen4 and Zen5 results are about even at the same PPT targets, which i would say is good thing (although the Z5 beat your Z4X3D at 120w example above)

According to AMD IPC is 23% better in Blender and we know that N4P is 28% more efficient than N5, so at 120W Zen 5 should be 25% faster than Zen4@120W.

 

[igor_kavinski]

so at 120W Zen 5 should be 25% faster than Zen4@120W.

But maybe 120W is not enough juice for the beefy Zen 5 cores?

 

[tsamolotoff]

If this continues, it will be interesting to see what happens at the big-end, although i suspect the X3D is drawing close to its limit

I personally believe Z5 can be simply too power starved at 120W (minus 20-25W or so of iodie), wider cores probably have to get more amperes than Z4 to clock properly. Well, I surely hope so, I was a bit hyped with linx results and this iso-power parity is not what I've expected.

 

[PJVol]

I've got an average score in Blender 3.3 for the Ryzen 9 7950X in Windows:
618
(cut several outliers below 500)

Loading…

docs.google.com

Blender - Open Data

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

 

[H433x0n]

According to AMD IPC is 23% better in Blender and we know that N4P is 28% more efficient than N5, so at 120W Zen 5 should be 25% faster than Zen4@120W.

According to Anandtech’s charts N4P is -22% power or +11% performance at isopower over base N5 V1.0. It's pretty unlikely that Zen 4 uses this same version of N5 when it launched 2 years later.

 

[LightningZ71]

Didn't AMD also use a somewhat customized version of N5 for Zen4?

 

[adroc_thurston]

Didn't AMD also use a somewhat customized version of N5 for Zen4?

They all do, DTCO's name of the game in this day and age.

 

[Abwx]

According to Anandtech’s charts N4P is -22% power or +11% performance at isopower over base N5 V1.0. It's pretty unlikely that Zen 4 uses this same version of N5 when it launched 2 years later.

22% less power mean 1/0.78 = 28% better perf/watt at isofrequency.

And a hint about the N4P process characteristic :

if power decrease by 0.78x at isofrequency, that is 1.28x less power, while frequency increase 1.11x at isowatt then it means that in the considered domain the Power/frequency curve is of the form :

P(f) = f^(ln1.28/ln1.11) = f^2.37

Overall in Blender Zen 5 would consume 23% more for 23% higher throughput at same frequency and N5 process.
Given that N4P consume 0.78x the power then Zen 5 should consume 1.23 x 0.78 = 0.96x the power at 23% higher througHput, this let just enough frtequency upllft to get 25% better perf at 1x the power.

But maybe 120W is not enough juice for the beefy Zen 5 cores?

If it needed more power at same perf they coudnt advertise higher perf, it would be at best same perf at higher power and the 9600X/9700X wouldnt be 65W, so perf/watt was increased by the process and the augmented IPC if it is used to slightly downclock the chip at still higher perf.