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

[DisEnchantment]

Speculate at will

 

[APU_Fusion]

Now we have Apple M4 with zen 5 being better and worse than it and zen 4 and zen 3 and lunar lake and arrow lake and compilers and the cache and cpu-ID gimmicks, ipc versus st versus mt versus e core versus c core, rage, sorrow, flaming, hatred, personal vendettas and a partridge in a pear tree. Chef’s kiss to thread

 

[Hitman928]

A quick Google said 2.75mm² vs 2.8mm².

Edit: but I'd defer to Hitman's numbers above

I won't take credit for the Zen 4 numbers, they were pulled from here, https://www.semianalysis.com/p/zen-4c-amds-response-to-hyperscale .

Edit: M2 numbers here, https://www.semianalysis.com/p/apple-m2-die-shot-and-architecture

 

[naukkis]

This looks like a single core holding 5.1 GHz takes 18 W - 20 W to me. The initial power rush is a transition and probably due to multiple cores spinning up at the beginning of the test as the scene is loaded but then settles down to 18 W - 20 W once the pure 1T compute starts.

Edit: This is also package power, so the core power is obviously less than that and scaling up to 5.7 GHz won't be nearly as dramatic as you make it seem.

Little better than totally ridiculous, but it also can seen that it cannot sustain 5.1Ghz but scales down to flat 5GHz. But that kind of power levels are not acceptable at all, they are straight desktop-class power uses. For comparison Intels problematic Raptor lakes only start have such a 1T power levels somewhere 13600K levels, even 14900K does not consume more than about 35W at full 1T boost.

 

[Hitman928]

Little better than totally ridiculous, but it also can seen that it cannot sustain 5.1Ghz but scales down to flat 5GHz.

So 33% less power than you were quoting is only a little better?

But that kind of power levels are not acceptable at all, they are straight desktop-class power uses.

MTL actually takes as much, if not more, power to reach its single core boost.

For comparison Intels problematic Raptor lakes only start have such a 1T power levels somewhere 13600K levels, even 14900K does not consume more than about 35W at full 1T boost.

You sure about that?

Seems like you are judging AMD by a very different standard than Intel.


Edit:

Found a comparison for Zen 4 (mobile) and MTL. Looks like they improved ST PPW significantly for STX. Do you still think 20 W is an unacceptable amount of power? This isn't even the 185H which has a 5.1 GHz boost, the 155H boosts to 4.8 GHz.

 

[Abwx]

even 14900K does not consume more than about 35W at full 1T boost.

For the full system 56W idling and 121W with Cinebench 1 T, the delta is 65W, wich amount to something like 50W just for the core excluding the uncore since it s already accounted in the idle power, that s almost 50% more power than your 35W wich is a made up number, guess that you dont even read reviews and are just throwing random guesses.

Intel Core i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks

i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks / Testmethodik / Multi-Core-Leistung / Single-Core-Leistung

www.computerbase.de

 

[Thunder 57]

So assuming those prices are real it looks like I won't be eating cat food. Also I can strikeout the last Zen 5 myth:

April launch
30%+ IPC
$499 9950X (Pending. 8 Ball says, "Outlook Hazy")

What is odd to me is that if it really is only a printing error that caused this 1-2 week delay, why not stick to the schedule and have reviews out today (or yesterday)? Are they really waiting to see what Intel does and hope that people bother re-benchmarking RPL? I think the story has become so large they will be bechmarked with the new microcode. Well, they say waiting is the hardest part.

EDIT: I will add for clarity that I am not trying to crap on AMD or anyone. More about being smarter about what we read and maybe use a little "Trust, but verify". Just thought I'd add that since this is a vendor thread and safe space.

 

[Thibsie]

Strix point seems worse and worse. It's ridiculous that it even cannot sustain max ST boost clocks on sub 30W devices - what full Zen5 @5,.7GHz consumes, full 100W @ ST workloads? Thats starts to be as ridiculous as Intel Raptor lake fiasco.

If Asus is clueless enough to sell a laptop with deficient cooling, how's that AMD's fault ?

 

[Abwx]

If Asus is clueless enough to sell a laptop with deficient cooling, how's that AMD's fault ?

The cooling is good enough, in ST Computerbase made a test with 300 seconds duration and there s no throttling.

In performance mode, wich is 30-33W, there s no throttling in MT, nnly if standard or whisper mode are enabled, seems that we have an urban legends that is forged live in this thread, lol...

AMD Ryzen AI 9 HX 370 „Strix Point“ im Zenbook S16 im Test

AMD „Strix Point“, die neue APU mit Zen-5(c)-Kernen, RDNA 3.5 und starker NPU, ist da. Im Test ist der Ryzen AI 9 HX 370 sehr effizient.

www.computerbase.de

 

[poke01]

So is A

The cooling is good enough, in ST Computerbase made a test with 300s seconds duration and there s no throttling.

In performance mode, wich is 30-33W, there s no throttling in MT unless standard or whisper mode are enabled, seems that we have an urban legends that is forged live in this thread, lol...

AMD Ryzen AI 9 HX 370 „Strix Point“ im Zenbook S16 im Test

AMD „Strix Point“, die neue APU mit Zen-5(c)-Kernen, RDNA 3.5 und starker NPU, ist da. Im Test ist der Ryzen AI 9 HX 370 sehr effizient.

www.computerbase.de

i wonder which laptop David is using then?

 

[Josh128]

What is odd to me is that if it really is only a printing error that caused this 1-2 week delay, why not stick to the schedule and have reviews out today (or yesterday)? Are they really waiting to see what Intel does and hope that people bother re-benchmarking RPL? I think the story has become so large they will be bechmarked with the new microcode. Well, they say waiting is the hardest part.

Not only that, but if the problem was with 9700X and some 9600X IHS labeling, why would that have any affect at all on 9900X and 9950X launch dates? If its all the delay was about, those should have launched without issue today. At the very least, they would have launched the same day as the "problem" SKUs-- but they launch a full week after that -- that doesnt make any sense whatsoever if the issue is just labeling of 9700X and 9600X.

 

[Abwx]

So is A

i wonder which laptop David is using then?

Dunno what he s using but at Computerbase they measure 18-19W average SoC power and the frequency is constant at about 5GHz according to their graph, eventually he got a pre production laptop that has not the final bios.

 

[Hitman928]

The cooling is good enough, in ST Computerbase made a test with 300 seconds duration and there s no throttling.

In performance mode, wich is 30-33W, there s no throttling in MT, nnly if standard or whisper mode are enabled, seems that we have an urban legends that is forged live in this thread, lol...

AMD Ryzen AI 9 HX 370 „Strix Point“ im Zenbook S16 im Test

AMD „Strix Point“, die neue APU mit Zen-5(c)-Kernen, RDNA 3.5 und starker NPU, ist da. Im Test ist der Ryzen AI 9 HX 370 sehr effizient.

www.computerbase.de

The ST test shows a very minor throttling to 5 GHz from 5.1 GHz. It's not a power throttle, though, as it is only using <= 20 W. It's got to be thermal throttling due to the hot spot effect and the laptop probably trying to keep fans low.

i wonder which laptop David is using then?

Pretty sure it's the same one, he was showing the same sustained frequency as CB.

Edit: Looking again, he posted a couple of different times, one with it throttling to 4.8 GHz. That one is probably the 13 inch model that has less cooling capacity than the 16 inch model that CB tested.

 

[Abwx]

The ST test shows a very minor throttling to 5 GHz from 5.1 GHz. It's not a power throttle, though, as it is only using <= 20 W. It's got to be thermal throttling due to the hot spot effect and the laptop probably trying to keep fans low.

Pretty sure it's the same one, he was showing the same sustained frequency as CB.

That s a very small throttling because at the end of the test it still get to roughly 5GHz, the rest of the time it s about the same with a short peak at 5140MHz.

All this is within error margins, unfortunately they didnt made a graph for the temp but since it doesnt throttle in MT there s no reason that temp could be too high in ST even accounting for the thermal density.

 

[poke01]

I don’t think ASUS is the right OEM to judge, I rather wait for Lenovo or HP make better comparisons.

Also isn’t it strange that all of not most of the Strix laptops available right now are from ASUS? Same was for the X Elite.

 

[HurleyBird]

The ST test shows a very minor throttling to 5 GHz from 5.1 GHz. It's not a power throttle, though, as it is only using <= 20 W. It's got to be thermal throttling due to the hot spot effect and the laptop probably trying to keep fans low.

If it's thermal throttling at all, is it possible that we're also seeing more aggressive process migration to move the hot spot around? That's not unheard of when mitigating core temp, whereas here it's likely skin temp, but maybe they're treated similarly.

 

[Hitman928]

That s a very smal throttling because at the end of the test it still get to roughly 5GHz, the rest of the time it s about the same with a mall peak a 5140MHz.

All this is within error margins, unfortunately they didnt made a graph for the temp but since it doesnt throttle in MT there s no reason that temp could be too high in ST even accounting for the thermal density.

You don't think that > 10 W in one core might cause thermal throttling because < 30 W across 12 cores doesn't thermal throttle? You realize that is a very large jump in power density right? It could easily be hot spot temp causing the throttling. . .

 

[Hitman928]

If it's thermal throttling at all, is it possible that we're also seeing more aggressive process migration to move the hot spot around? That's not unheard of when mitigating core temp, whereas here it's likely skin temp, but maybe they're treated similarly.

I don't think skin temp is the problem, that's more a problem when pushing total power. With the large difference in ST to MT clocks, the hot spot temperature is much more likely the culprit. Asus could probably prevent the throttling but they don't want the fans kicking up too much every time the CPU boosts on 1 or 2 cores.

 

[HurleyBird]

I don't think skin temp is the problem, that's more a problem when pushing total power. With the large difference in ST to MT clocks, the hot spot temperature is much more likely the culprit. Asus could probably prevent the throttling but they don't want the fans kicking up too much every time the CPU boosts on 1 or 2 cores.

If that's the case, then more aggressive process migration probably accounts for the remainder of the performance deficit compared to expectations that isn't accounted for by the relatively minor decrease in average clocks. Otoh, some review samples may see a greater reduction to clocks than others.

 

[Abwx]

I don’t think ASUS is the right OEM to judge, I rather wait for Lenovo or HP make better comparisons.

Also isn’t it strange that all of not most of the Strix laptops available right now are from ASUS? Same was for the X Elite.

They have one month or so exclusivity, previously it was HP and then Lenovo.

You don't think that > 10 W in one core might cause thermal throttling because < 30 W across 12 cores doesn't thermal throttle? You realize that is a very large jump in power density right? It could easily be hot spot temp causing the throttling. . .

At the 50s mark it hit 4996MHz but at the 190s mark it goes at 5105MHz, so if there s thermal density motivated throttling it couldnt get back to 5105 after 190s, the frequency would just gradually and lightly tank.

More likely that AMD s boost algorithm is very fast and if a slightly lower loading occur for a very short duration the core frequency will be reduced very shortly before the max loadin will get it increasing again in a matter of tens of microseconds, btw ST test was done in perf mode wich mean with the fan spinning faster than in standard mode.

 

[Hitman928]

If that's the case, then more aggressive process migration probably accounts for the remainder of the performance deficit compared to expectations that isn't accounted for by the relatively minor decrease in average clocks.

Maybe someone will investigate the thread migration. However, I don't think there is an issue with the ST performance deficit. The GB results we have seem to be fine (very short tests with breaks in between so thermal issues don't crop up). The main issue was the Anandtech results, but we don't know what frequency their laptop ran at. Their SPEC run is obviously throttling more than 100 MHz though.

 

[Hitman928]

At the 50s mark it hit 4996MHz but at the 190s mark it goes at 5105MHz, so if there s thermal density motivated throttling it couldnt get back to 5105 after 190s, the frequency would just gradually and lightly tank.

More likely that AMD s boost algorithm is very fast and if a slightly lower loading occur for a very short duration the core frequency will be reduced very shortly before the max loadin will get it increasing again in a matter of tens of micrseconds, btw ST test was done in perf mode wich mean with the fan spinning faster than in standard mode.

Nah, the clock will always bounce around a bit, especially if thermally throttled, because the load isn't actually constant. Different gates and different number of gates are firing from moment to moment which will briefly change the severity of the hot spot effect. It'd be interesting to see how much core migration is happening as well.

Edit: Huang said he plans on using some sort of air conditioning on the laptop (i.e., using cooled air) to stop the throttling, so . . .

https://twitter.com/x/status/1818689048921489702

"I plan to find a way to suppress it with air conditioning this weekend"

 

[poke01]

Apple's core sizes are decently bigger than Zen cores on equivalent nodes with similar design frequencies. AMD's high performance cores come close to reach high frequencies.

Edit: I broke it down here but I'll copy again if you don't want to read the full post:

Zen 4 Core = 2.56 mm2 with max boost of ~5.7 GHz.
Zen 4c Core = 1.43 mm2 with max boost of ~3.7 GHz.
M2 core = 2.76 mm2 with max boost of 3.5 GHz.

Zen 4 Core + L2 = 3.84 mm2 with max boost of ~5.7 GHz.
Zen 4c Core + L2 = 2.48 mm2 with max boost of ~3.7 GHz
M2 Core + L2 ~ 7.06 mm2 with max boost of 3.5 GHz

So Apple is more cache to make up for the frequency difference and AMD using higher frequency but using less cache?

Obviously there is a power consumption difference in ST when boosting to 5.7GHz vs 3.5GHz. It’s all design trade off, AMD can spam more cores (useful for server) because they use less area and Apple can reach excellent 1t performance in a phone or tablet due to their design. I see why each approach makes sense.

It would be interesting to compare Intel cores as well once they get to a proper node.

 

[Kryohi]

You don't think that > 10 W in one core might cause thermal throttling because < 30 W across 12 cores doesn't thermal throttle? You realize that is a very large jump in power density right? It could easily be hot spot temp causing the throttling. . .

Wait, doesn't Windows rotate single thread workloads among cores?
On Linux I don't think I ever encountered an issue like this, every 10 seconds or so a thread is passed to a different core, at least according to my system monitor.
Perhaps the Asus laptop is doing something like this but not all cores can reach 5.1GHz?

Edit: I'm on vacation off-country since Sunday and got a serious case of food poisoning after 2 days, but I must say that the last 15 pages of this thread managed really well to keep me distracted and entertained.
One thing that keeps being repeated though is that the high cross-CCX latency between the zen 5 and zen 5C cores is hurting performance. May I ask if there is any proof of that in the actual (non-synthetic) benchmarks ran so far? From the Phoronix review nothing like that had emerged afaik.

 

[Hitman928]

So Apple is more cache to make up for the frequency difference and AMD using higher frequency but using less cache?

Apple caches are big, yes. Helps IPC and efficiency. Apple uses a shared L2 for the P-core clusters, though, and no L3 (but they have access to the sizeable system cache) so a direct comparison past L1 cache is tricky. Whatever the magnitude is, Apple does use more sizeable cores with lots of cache compared to AMD who goes higher frequency.

Obviously there is a power consumption difference in ST when boosting to 5.7GHz vs 3.5GHz. It’s all design trade off, AMD can spam more cores (useful for server) because they use less area and Apple can reach excellent 1t performance in a phone or tablet due to their design. I see why each approach makes sense.

Yes, this is part of the divergence in design approaches. AMD needs a core that can scale to > 100 cores, which means the cores themselves can't be too big. Their dense cores especially are quite good in terms of PPA since they don't chase high frequency and so they can scale all the way to 192 cores on a leading process. Zen 4 introduced dense cores and Zen 5 is the first time they are actually changing parts of the inner core between product lines as well (FP width changes). I expect Zen 6 will see even a bit more specialization between product lines.

It would be interesting to compare Intel cores as well once they get to a proper node.

P cores on LNL (TSMC N3B) are ~4.25 mm^2 with L2 cache.

 

[Hitman928]

Wait, doesn't Windows rotate single thread workloads among cores?
On Linux I don't think I ever encountered an issue like this, every 10 seconds or so a thread is passed to a different core, at least according to my system monitor.
Perhaps the Asus laptop is doing something like this but not all cores can reach 5.1GHz?

As far as I know, on AMD systems there are cores that are marked as preferred that can reach max boost threads and the scheduler is supposed to keep ST loads on preferred cores only, but I'm not 100% on this. I know that other laptops don't seem to show this throttling problem, so I don't think it's a core migration issue.