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

[DisEnchantment]

Speculate at will

 

[Josh128]

Here we are getting excited to even see a single actual +16% uplift vs 7950X, when Zen 5 is supposed to average +16% vs Zen 4 in IPC uplift-- the sad reality is that it seems to fall quite a bit short of that. When looking at absolute performance we've already seen it definitely falls well short of +16% average vs Zen 4, SKU vs SKU.

 

[igor_kavinski]

Don't worry. Deep down somewhere in a basement miles below Earth, there is a bunch of pale AMD engineers tweaking every little parameter to finally figure out the right combination to hit 35% IPC

I would give them a year, at the very least. It's hard work, this firmware business.

 

[CouncilorIrissa]

Here we are getting excited to even see a single actual +16% uplift vs 7950X, when Zen 5 is supposed to average +16% vs Zen 4 in IPC uplift-- the sad reality is that it seems to fall quite a bit short of that. When looking at absolute performance we've already seen it definitely falls well short of +16% average vs Zen 4, SKU vs SKU.

GB6 pretty consistently shows 15-17% uplift for both Strix and GNR tbh. Too bad it's mostly dragged up by SIMD-heavy tests.

Gigabyte Technology Co., Ltd. B650E AORUS PRO X USB4 vs Gigabyte Technology Co., Ltd. X670E AORUS XTREME - Geekbench

browser.geekbench.com

Seems like it will take a 9950X3D to beat this annoying 7950X in every subtest.

Zen 5 has been consistently weak in Navigation subtest, you're not beating Zen 4 in this particular subtest iso-clock.

 

[igor_kavinski]

Zen 5 has been consistently weak in Navigation subtest, you're not beating Zen 4 in this particular subtest iso-clock.

Yeah. Even with higher frequency, it's gonna be hard for Zen 5 but I'm rooting for 9950X3D to manage it.

 

[biostud]

After seeing the efficiency graphs @ techpowerup I have less hope for higher clocked 3D cache version of zen5 over zen4, unless the redesigned core somehow can dissipate heat better.

AMD Ryzen 7 9700X Review - The Magic of Zen 5

The AMD Ryzen 7 9700X boasts a robust eight-core design and exceptional single-threaded performance with great power efficiency. In our review we saw good gains with PBO overclocking and TDP increases. Unfortunately, the $360 price tag is steep, and there are several good alternatives available.

www.techpowerup.com

Oh, and I don't regret buying a 7800X3D

 

[Doug S]

Didn't Geekbench pause and start its loads? I think that was exactly the scenario where Intel could be affected because the VID cap fix could affect the speed of boost ramping. Continuous 1T load could be unaffected, bursty loads affected. Also, if TVB is fixed to actually not allow the highest clocks over 70C temperature as it is supposed to be, then that could worsen Intel's 1T score. The effect of Inte default settings is a separate thing, actual 253W PL2 likely only affects MT score, but the settings were touching other things too...

Yes, this is exactly the case. Intel's explanation was that the CPU tried to anticipate for voltage droop when suddenly demanding a lot more power (i.e. boosting a core from idle through its stages to turbo in quick succession, or especially boosting a bunch of cores all at once) so they gave it more voltage than what was needed/desired to precompensate for voltage droop.

It isn't clear whether they are trying to compensate for voltage droop from the integrated voltage regulators ("FIVR") or from the motherboard regulators. I would think it would have to be the latter, since they don't control those, and there's some variation between the voltage regulation quality on cheap vs high end motherboards. If that's the case, then ironically it may be the case that your CPU was MORE likely to be damaged if you used a high end board with the best quality voltage regulators, since those would have less voltage droop which would have lead to Intel's droop compensation algorithm overshooting the desired voltage!

 

[SK10H]

After seeing the efficiency graphs @ techpowerup I have less hope for higher clocked 3D cache version of zen5 over zen4, unless the redesigned core somehow can dissipate heat better.

AMD Ryzen 7 9700X Review - The Magic of Zen 5

The AMD Ryzen 7 9700X boasts a robust eight-core design and exceptional single-threaded performance with great power efficiency. In our review we saw good gains with PBO overclocking and TDP increases. Unfortunately, the $360 price tag is steep, and there are several good alternatives available.

www.techpowerup.com

Oh, and I don't regret buying a 7800X3D

Seriously that entire page has no predictive value on the Zen5 x3d boast clock as it's not temp/power being the limit.

I don't regret swapping to a 7800x3d after the first reliable hint of Zen5 train wreck. 😛 I haven't bother reading any review yet too 🤣, just going to rely on forum warriors tweaking graph as most reviewers are just not capable to show the real meat I want to see due to knowledge, time constraint or just flat out laziness.
That said, not going to upgrade until Zen6 with the new uncore, first time skipping a Zen generation since Zen1. Even had the Zen+ replacement for a swap in refresh.

 

[StefanR5R]

I wish the tech press would get more into the mindset they had 2018/2019, where people were tuning the nuts off of their rigs and analyzing things like overclocking and ram to the finest detail.

Maybe some will do it at some point after the launch. But not at launch. They receive samples very shortly before the review publishing embargo ends, and after they got the samples, they continue to receive one wonky beta BIOS after another. Yet on the other hand, there is that certain pressure to have the report ready immediately when the embargo ends. My impression is this has gotten worse over time.

Besides the race to publish immediately after the embargo falls, there is also the issue that they have to return the samples. But I don't know how quick. (There were even launches in the recent past, at which one and the same sample had to make the rounds through several reviewers because for whatever reason the number of samples per continent was extremely limited.)

 

[Jan Olšan]

Here we are getting excited to even see a single actual +16% uplift vs 7950X, when Zen 5 is supposed to average +16% vs Zen 4 in IPC uplift-- the sad reality is that it seems to fall quite a bit short of that. When looking at absolute performance we've already seen it definitely falls well short of +16% average vs Zen 4, SKU vs SKU.

Technically the Geekbench thing is a combo of several tests so it is a bit better than just "single actual uplift", that result is an average of a few. Similarly SPEC (which is SPECINT + SPECFPU and those both have several subtests that are averaged into the scores). Of course, they are still more aiming to be approximations of real software than being real software themselves.

 

[Josh128]

Technically the Geekbench thing is a combo of several tests so it is a bit better than just "single actual uplift", that result is an average of a few. Similarly SPEC (which is SPECINT + SPECFPU and those both have several subtests that are averaged into the scores). Of course, they are still more aiming to be approximations of real software than being real software themselves.

Of course. But lets be real- when AMD announced +16% average IPC and +17% Cinebench R23 perf, nobody expected LESS THAN +14% and 10% for R23 ST and MT actual performance uplifts, SKU vs SKU. I certainly didnt. Even with PBO it doesnt come close to +17% R23 perf. As it stands, 9950X will score less than 14900KS in ST R23.

When 16% is the average, there should be just as many results over 16% as there is under 16%. Sadly, it is just not the case.

 

[desrever]

When 16% is the average, there should be just as many results over 16% as there is under 16%. Sadly, it is just not the case.

There are way more over 16% than under 16% based on application performance.

 

[Abwx]

Of course. But lets be real- when AMD announced +16% average IPC and +17% Cinebench R23 perf, nobody expected LESS THAN +14% and 10% for R23 ST and MT actual performance uplifts, SKU vs SKU. I certainly didnt. Even with PBO it doesnt come close to +17% R23 perf. As it stands, 9950X will score less than 14900KS in ST R23.

When 16% is the average, there should be just as many results over 16% as there is under 16%. Sadly, it is just not the case.

7700X run at 5.28Ghz in Cinebench 2024 while the stock 9700X run at 4.35GHz and is still 2% faster.

528/435 x 1.02 = 23.8% better IPC.

AMD Ryzen 5 9600X und Ryzen 7 9700X im Test: IPC, Caches, RAM, AVX-512, TDP, Kernverbrauch und mehr

Ryzen 5 9600X & Ryzen 7 9700X im Test: IPC, Caches, RAM, AVX-512, TDP, Kernverbrauch und mehr / AMD Ryzen 7 9700X mit 65 Watt TDP im Fokus

www.computerbase.de

 

[Markfw]

There are way more over 16% than under 16% based on application performance.

this depends on what applications you run. Again, it may not equal what AMD used. Example: ifs all but proven based on SOME benchmarks that the avx-512 performance is twice what it was on Zen 4. Most people don't care about it, but I am actually buying one JUST for that purpose. I want to see phoronix, and especially if they include good benchmarks on that.

 

[Abwx]

TSMC s N4P provide 28% better perf/watt than their regular N5 but it happen than Zen 4 use N5P wich has 11% better perf/Watt than the vanilla N5.

Hence the perf/Watt improvement at isofrequency brought by N4P is only 1.28/1.11 = 15.3% better perf/Watt or 5.7% better perf at isopower.

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

www.anandtech.com

 

[Josh128]

7700X run at 5.28Ghz in Cinebench 2024 while the stock 9700X run at 4.35GHz and is still 2% faster.

528/435 x 1.02 = 23.8% better IPC.

AMD Ryzen 5 9600X und Ryzen 7 9700X im Test: IPC, Caches, RAM, AVX-512, TDP, Kernverbrauch und mehr

Ryzen 5 9600X & Ryzen 7 9700X im Test: IPC, Caches, RAM, AVX-512, TDP, Kernverbrauch und mehr / AMD Ryzen 7 9700X mit 65 Watt TDP im Fokus

www.computerbase.de

I never said one cant arrive at the IPC number. I said perf. Without manual OC, its literally impossible to actually get +17% ST R23 perf on 9700X vs 7700X.

 

[yottabit]

When 16% is the average, there should be just as many results over 16% as there is under 16%. Sadly, it is just not the case.

that would technically be median not average (mean)

but putting the smarty pants aside I concur with the feels of your argument

 

[Abwx]

I never said one cant arrive at the IPC number. I said perf. Without manual OC, its literally impossible to actually get +17% ST R23 perf on 9700X vs 7700X.

For 17% better perf at same frequency you ll have to increase power by 17% if uarch efficency and process are the same.
There s not enough headroom with N4 to increase perfs by more than 11% at a same power, and that s assuming that Zen 5 has same uarchitectural efficency than Zen 4, wich is unlikely, if Zen 5 has such efficency being 5% lower then the headroom is about 5%.

Edit : With N3E they would had gained either 18% higher perf/watt ( = 15% lower power at isofrequency) or 7% higher perf at same power, all those numbers in respect of N4P.

 

[H433x0n]

Gamers Nexus did a video on Zen 5 power efficiency.

TLDR; 9700X is slightly less efficient than its Zen 4 counterparts in gaming and compression/decompression workloads.

However, there are perf/watt improvements in blender at default TDP (roughly ~10%).

 

[Abwx]

Gamers Nexus did a video on Zen 5 power efficiency.

TLDR; 9700X is slightly less efficient than its Zen 4 counterparts in gaming and compression/decompression workloads.

However, there are perf/watt improvements in blender at default TDP (roughly ~10%).

He confuse perf and efficency.

What he s saying is that car A travel 100km at 111km/H within 0.9H all while using say 11 litres to do so.

And he compare it to car B wich travel 100km at 100km/H within 1H and 10 litres to do so.

That s performance, not efficency.

Efficency is to measure the distance that is traveled with 10 litres at a same speed, or wich is the same, the quantity of gasoline necessary to travel 100km at 100km/H.

 

[H433x0n]

He confuse perf and efficency.

What he s saying is that car A travel 100km at 111km/H ithin 0.9H all while using say 11 litres to do so.

And he compare it to car B wich travel 100km at 100km/H within 1H and 10 litres to do so.

That s performance, not efficency.

Efficency is to measure the distance that is traveled with 10 litres, or wich is the same, the quaNtity of gasoline necessary to travel 100km at 100km/H.

I’m guessing in your hypothetical, fuel is power measured in watts. If that’s the case, you often state it the other way for non gaming workloads.

 

[Ranulf]

From GN: FinalFantasy 14 Benchmark, power draw, 7600X vs 9600x. The spikes in the graph are main loading points for the benchmark. So from this data, the 9600X is pulling more power running the benchmark except at loading times.



Edit to add Power cost per year chart:

 

[Abwx]

I’m guessing in your hypothetical, fuel is power measured in watts. If that’s the case, you often state it the other way for non gaming workloads.

In watt/hours, that s the comsumption but this doesnt say what is the efficency since the latter is speed dependant, when you increase speed by a factor X then efficency is reduced by the same factor, it work the same for CPUs and cars, CPUs frequency increase, at best, as a root square of energy and cars speed increase as a root square of kinetic energy.

As you surely deducted it you can trade a better efficency for a higher speed, if you have a CPU that is 21% more efficient at same work within a same time you can trade this efficency for 10% more work performed in a same time, or wich is the same, to do a same work within a 1.1x lower time.

So far for CPUs you cant just increase efficency because you have to account for the competitive landscape, so generaly the choice is to trade part of the augmented efficency for a little more perf by using the same TDP as the previous product.

 

[DaaQ]

AMD didn't make the 9950X to excel at Geekbench. I think AMD knows pretty well that it will be hard to get existing AM5 users to upgrade. They are most likely hoping for AM4 stragglers with 5900X/5950X to make the jump to Zen 5. And now add to that thousands of Intel users who will be wary of Core series CPUs after the Raptor Lake fiasco.

Upgrading from AM3+ to AM5, will be good performance jump. Patiently waiting on X3D info.

 

[cebri1]

Gamers Nexus video addressing all the most common claims around “Reviewers do not know how to test”

 

[eek2121]

this depends on what applications you run. Again, it may not equal what AMD used. Example: ifs all but proven based on SOME benchmarks that the avx-512 performance is twice what it was on Zen 4. Most people don't care about it, but I am actually buying one JUST for that purpose. I want to see phoronix, and especially if they include good benchmarks on that.

People also underestimate the usefulness of AVX-512. While it is not the easiest to add in support for, it provides fantastic speedups for the relevant workloads. With Zen 4 and Zen 5 becoming mainstream, developers have more incentive to add support for it. Especially as Intel continues to falter.