AusWolfHave you read a single thing I said? It's not about power! It's about heat dissipation! Why do you think some mobile SoCs run hot even though they consume around the 5 W range? Jesus...

The R5 3600 was a problem not because it was hungry, but because it couldn't dissipate its heat in a small case with limited airflow, while the i7 11700 can. The 3600 got to 90 °C at around 80 W, which is not even the default power limit (it's 88 W - read about PPT). The 11700 needed 130 W to reach that temperature with the same cooler.

If you still don't get it, I don't know how else you would.

But the temperature makes no difference, especially in your comparison. Yes, the 3600 reached a higher core temperature. But it still maxed out the power, so clearly it was being cooled enough. It might even have a higher core temperature at even lower wattage, because of how temperature gradient works on the core. Seriously, Celsius does not matter unless it actually throttles, which it doesn't seem like it did, you just see a number and go, ooo number big, competitor number small, must be better.

tabascosauzDisabling CPB is AMD's equivalent of disabling Turbo Boost and running exclusively at base clock. It's just not so apparent when base clock is as high as it is for Zen 4 (4.7GHz).

Not sure where this silver bullet idea comes from. Literally nobody does this. You don't buy a 12700K to run it at 3.6GHz either.

All I see here is that the aggressive power limiting and undervolting practices for 5800X and 5800X3D are about to become the inevitable norm, for all Zen 4 owners.

AMD is clearly pushing the envelope of N5 hard. 4.7GHz for 1681pt is an appreciable step up in IPC but not much at all over Zen 3 (~4.85GHz for 1600pt), although the temps and power are an impressive demonstration of V-F curve on N5 (I/O die and Fabric still take a chunk of power, out of the Package Power). Doesn't sound like AGESA is ready yet.

They still are relying on pushing clockspeed to get the performance they want. Every time they've done this, the launch processors don't clock up to expectations. Hopefully not the case this time.

I'd go as far as saying the AGESA for Zen 3 isn't ready yet... the blasted EDC bug has been grilling my rig hard

After thinking a while about the posted CBr23 screenshots, we don't even know what the ambient temps were. Cooling solution could the the most crappy Dell stock cooler for all we know, which can barely handle 95W.

But the the 92.1°C temp. I think it must be the single core CBr23 run temp. But it could also be the temp from a Prime 95 single core run, which gets every CPU very hot since you just hit one core with insane loads.
The 110W are likely from the all core Cinebench run. And the temperature at all core load is usually way lower then the single core run temperature. The conclusion is drawn from my Zen 2 and Zen 3 CPUs. Even the R5 3600 behaves in that way.
The 5.45GHz boost could even be the idling boost when CBr23 switches from the multicore test to the single core test. We know these are all the max values, so no reset has been done on HWinfo between the single core and multi core runs.
The score are especially unreliable, since there is variance. HWinfo takes away quite some points from the Single core result, and that is assuming the Windows scheduler didn't throw around the single core load on all cores, which can still happen.
Multi core should be more reliable, but still only one score we see.

@AusWolf I get your concern about the high temperature. But keep in mind, it is most likely under heavy single core load. All core should be much lower. The CPUs are also designed to run at these temps, others already mentioned the heat density you get with these tiny Zen4 cores compared to the relatively large Alder Lake cores.

@phanbuey could you please post the power draw of you 12600K to your Cinebench R23 scores? Not really relevant as comparison, since we don't know anything about cooling, roomtemp or silly stuff running in the background of the 7600X. But i am curious :)

AM4isGODMine is on water temp as i have a temp sensor in the loop with a USB aqua computer device.

Remember i have 6 120mm fans on two 360 radiators, so they can run very slow and quiet without really compromising the cooling.

Also my GPU is in the loop too, less than 50c gaming, which is very nice.

AusWolfI wonder what cooler they used. I don't like that 92.1 °C with only 110 W one bit. :wtf: This is why I hated the R5 3600 that I gave away after a week or so. Either the IHS design hasn't improved despite the switch to LGA, or chiplets will always run hot. :(

110w is easy to cool on ambient

what can be improved with change from PGA to LGA? its just copper block. Zen is manufactured on smaller node, so its denser => harder to cool

What cooling is needed for 60.2 W peak?
13003 is mediocre for 4.7GHz which is the base clock and not stressing out the silicon at all.
If you stress a low end 5600 non-X at 4.6GHz it achieves around 11990 so the IPC gain is around 6.2% not 9%.
I'm not suggesting that the IPC claim that AMD made was wrong just that it doesn't substantiated in this result and maybe with a newer BIOS will show a little bit better performance.
Still even if with a newer BIOS achieves the 9% IPC increase, 13600K/KF will be around +50% faster in apps that scale well with core count like some rendering or media encoding apps.

For IPC claims you need to have fixed clocks, none of the leaked benchmarks are really showing "IPC" difference.

7600X (105W TDP/142W PPT) has 4.7GHz base which means if the cooling is adequate it achieves at least 4.7GHz in 99.9% of apps.
If the tester didn't change something in the settings the processor logically would be hitting at least the advertised base in an app like CB23 since it had 56 °C and 60.2W peak and the cooling system was a dual-fan AIO kit...
So at best case 6.2% IPC increase.

"According to the HWINFO output, the CPU runs relatively cool at 56 °C and 60.2W peak, while core clock is somewhere in 4.7 GHz range"

RandallFlaggAnd more on topic -

https://twitter.com/i/web/status/1568231095145058304

That score is incredibly impressive, but it also is a golden sample 12900K that went through an extreme amount of fine tuning and memory tuning.
This is as close as you can reasonably get to the theoretical performance per Watt ceiling of Alder Lake's P-cores in a desktop environment.
Yet it still doesn't even begin to touch the performance per watt efficiency of the 7600X's 60 W score.

XeanoaThat score is incredibly impressive, but it also is a golden sample 12900K that went through an extreme amount of fine tuning and memory tuning.
This is as close as you can reasonably get to the theoretical performance per Watt ceiling of Alder Lake's P-cores in a desktop environment.
Yet it still doesn't even begin to touch the performance per watt efficiency of the 7600X's 60 W score.

No, not really, my pretty mediocre 12900k can get 12700 score at 54.1 watts, beating the 7600x handily in efficiency.

Efficiency SHOULD be compared watt normalized btw, you can't say X cpu at 50 watts is more efficient than Y cpu at 200 watts.

fevgatosNo, not really, my pretty mediocre 12900k can get 12700 score at 54.1 watts, beating the 7600x handily in efficiency.

Efficiency SHOULD be compared watt normalized btw, you can't say X cpu at 50 watts is more efficient than Y cpu at 200 watts.

Madness' 12900K was using only 6 P-cores, all other cores were disabled.
Efficiency was compared at the same core/thread count between finely tuned Alder Lake P-cores and the Zen 4 cores in the OP.

XeanoaMadness' 12900K was using only 6 P-cores, all other cores were disabled.
Efficiency was compared at the same core/thread count between finely tuned Alder Lake P-cores and the Zen 4 cores in the OP.

Yeah, im talking about 6/12 as well. 6/12 GC cores at 54w score 12700.

AusWolfI like small form factor builds. If a CPU runs hot on a test bench, or in a tower case with adequate airflow, then it will definitely throttle in a SFF system.

I wouldn't be so drastic. We don't know the pc specs of the guy who took those screenshots and what cooler was used. We don't even know if those scores are real or fake.

That said, a cpu can go near its tjmax only in rendering-like tasks, where all threads are pushed to 100% for several minutes, and only if the cpu cooler is not good for the job, or if it's overclocked with extreme voltage. But, who would use a +5GHz 12threads cpu for rendering in a shoe-box case?
Assuming you would buy that cpu, you would pair it with a proper air or AIO cooler and in gaming, or moving millions poligons in ZBrush, you won't ever come close to its tjmax, even inside a small case. So throttling won't be a problem for sure for a regular user (gaming, movies, graphics, internet, CAD, etc.).

All AMD cpus, at stock speeds can be undervolted by a good margin, and can be overclocked and undervolted at the same time. So if someone wants to use a so powerful cpu in a SFF closed case, he can tune it to whatever he needs and even use it as a workstation, assuming that he knows where to put his hands.

RandallFlaggThey made the surface area of their IHS too small.

The amount of heat you can x-fer, all else being equal, is going to be directly proportional to the surface area you have to transfer it through. I couldn't find the actual numbers on a quick search, but from visuals Intel's LGA 1700 chips have probably 1/3 to 1/2 more surface area for the IHS. This means they can transfer 1/3 to 1/2 more 'power' as heat provided the cooling solution can dissipate it.

That is probably the best explanation there is. No wonder the cooler's heatsink was cold to the touch on the 3600 even after several remounts, while it's reasonably warm on the 11700. Intel's IHS can transfer heat a lot better. I was hoping AMD would fix this for AM5, but it doesn't look like it so far.

vmarvI wouldn't be so drastic. We don't know the pc specs of the guy who took those screenshots and what cooler was used. We don't even know if those scores are real or fake.

That said, a cpu can go near its tjmax only in rendering-like tasks, where all threads are pushed to 100% for several minutes, and only if the cpu cooler is not good for the job, or if it's overclocked with extreme voltage. But, who would use a +5GHz 12threads cpu for rendering in a shoe-box case?
Assuming you would buy that cpu, you would pair it with a proper air or AIO cooler and in gaming, or moving millions poligons in ZBrush, you won't ever come close to its tjmax, even inside a small case. So throttling won't be a problem for sure for a regular user (gaming, movies, graphics, internet, CAD, etc.).

All AMD cpus, at stock speeds can be undervolted by a good margin, and can be overclocked and undervolted at the same time. So if someone wants to use a so powerful cpu in a SFF closed case, he can tune it to whatever he needs and even use it as a workstation, assuming that he knows where to put his hands.

That's the thing... I'm sure Zen 4 will do well in a standard chassis with a "proper cooler" as you said. I also know you can de-tune any CPU if you want to use it in a small form factor case with low airflow. The question is, by how much. My problem with the 3600 was that it couldn't even keep its factory default 88 W PPT without reaching throttling temps in my scenario (link in my signature), while as with the 11700, I could increase PL1 up to around 120-130 Watts.

Oh, and of course I don't want to render with a SFF PC. :) I only use Cinebench as a test to make sure it can withstand every reasonable task in all circumstances.

Hi,
I just call it a heat spreader with solder which is why once both are removed from in the way bare die sees a good drop in temps screw z height :cool:

AusWolfThat is probably the best explanation there is. No wonder the cooler's heatsink was cold to the touch on the 3600 even after several remounts, while it's reasonably warm on the 11700. Intel's IHS can transfer heat a lot better. I was hoping AMD would fix this for AM5, but it doesn't look like it so far.


That's the thing... I'm sure Zen 4 will do well in a standard chassis with a "proper cooler" as you said. I also know you can de-tune any CPU if you want to use it in a small form factor case with low airflow. The question is, by how much. My problem with the 3600 was that it couldn't even keep its factory default 88 W PPT without reaching throttling temps in my scenario (link in my signature), while as with the 11700, I could increase PL1 up to around 120-130 Watts.

Oh, and of course I don't want to render with a SFF PC. :) I only use Cinebench as a test to make sure it can withstand every reasonable task in all circumstances.

they cant fix it .. only if they used more dies with more dummy cores... if chiplet is small, you can have ihs, but the heat ll get transfered to an extend just through hot spot like on image :)

NaitoBecause voltage is a one size fits all approach when the chips leave the factory. You may be lucky enough to acquire a quality bin for which a lower voltage will not only potentially boost performance, but drop temperatures

I dont see binning making that big of a difference with heat, but higher mhz. The temps will still get high but you will have longer sustained boost with better chips. The longer boost will keep temps higher unless you lower frequency.

Leshythey cant fix it .. only if they used more dies with more dummy cores... if chiplet is small, you can have ihs, but the heat ll get transfered to an extend just through hot spot like on image :)

Then maybe as good the chiplet design is for yields, it's just as bad for heat transfer?

AusWolfThen maybe as good the chiplet design is for yields, it's just as bad for heat transfer?

yes .. smaller area cost less to manufacture, but also smaller area can transfer less heat :)

Leshyyes .. smaller area cost less to manufacture, but also smaller area can transfer less heat :)

Then I really don't understand why users praise chiplets so much, considering that the end product costs the same as traditional, monolithic CPUs. It's only AMD's gain, not ours.

RandallFlaggThat's flatly false. Surface area matters. The only other thing they can control is the materials which affects the heat transfer coefficient.

Think about the logical conclusion of what you just said. Can you transfer just as much heat through 1 square millimeter as 100 square millimeters?

Well yes you can - all else being equal, the temperature difference between the two surfaces would need to be 100 times greater to transfer the same heat.

Q = h A ΔT

Q = the rate of heat transfer

h = convection heat transfer coefficient

A = the exposed surface area, and

ΔT = the difference in temperature

You're misapplying your logic here. The surface area of the IHS is of ... well, at least tertiary importance, to its thickness and the size of the heat generating area of the die (assuming good-to-perfect core-to-IHS transfer etc). Heat is transferred most efficiently at high thermal deltas, meaning that as heat dissipates out through the IHS, it does a gradually worse job of moving further into the heatsink, as temperatures drop rapidly as you move out from the spot on the IHS directly above the core. The area directly above the core is always the most crucially important for cooling the core, as that is where the vast majority of thermal transfer into the cooler will take place. The rest of the IHS obviously helps, but increasing its size would see very low returns as the distance from those parts of the IHS to the core would increase. Remember: assuming a similarly sized heat source and the same thermal transfer between the two, your 1 square millimeter IHS will be much hotter than your 100 square millimeter one. And, of course, that example is ridiculous - we wouldn't be talking something like a 100x size increase, but maybe going from ~40x40mm (1600 mm²) to 45x45 (2025mm²) or maybe 50x50 (2500mm²). Those are 27% and 56% increases, all of which would be relatively far from the core, as the cores already have plenty of coverage from the current IHS design. So, while a larger IHS would likely have some beneficial effects, overall it would be very small as long as the heat source remained the same. You would need the IHS to be much, much larger - and also thicker, to aid in the IHS heating up more evenly - for this to have a meaningful effect. But even then it would come into conflict with other factors that would most likely diminish this effect significantly, such as thicker IHSes generally doing a worse job of transferring heat to the cooler - you want the thermal energy to have as short a path as possible to the cooler, after all.

The reason why Zen3 (and to some extent Zen2, and likely Zen4) runs hot is not because its IHS is too small, but because its core is very small, and its thermal density is thus very high. This means the IHS struggles more to effectively transfer said heat to the heatsink, but a larger IHS wouldn't alleviate that in any meaningful way. A thinner IHS could (though too thin and it becomes less efficient at horizontal thermal transfer, counteracting this effect somewhat), as could improved die-to-IHS bonds - or direct-die cooling.

Heck, this is the reason why cooling a 300W GPU is relatively simple - and can be done with even a 120mm AIO and a reasonably powerful fan - while cooling a 300W CPU is near impossible: GPUs spread their heat evenly across a large die and don't have IHSes, while CPUs have very small cores and have IHSes to protect them.

The only reasonable way of improving this is to improve the IHS materials. I've been speculating for years already if we'll ever see vapor chamber IHSes, and IMO that's not unlikely the way thermal density is going. Obviously it won't be trivial to make such a thing in a way that would survive the mounting pressure of a cooler on top of it, but it would be doable. And, of course, it would significantly increase BOM costs for CPUs.

(All of this is also of course dependent on the cold plate design of the cooler - an IHS is essentially necessary on any direct heatpipe contact HSF, as otherwise you risk the heat generating cores contacting just one heatpipe, which would lead to terrible cooling overall. On the other hand, a HSF with a soldered cold plate, or a water cooler with a finned cold plate, would cool far better without an IHS than with it.)

AusWolfThen I really don't understand why users praise chiplets so much, considering that the end product costs the same as traditional, monolithic CPUs. It's only AMD's gain, not ours.

It's only due to chiplets that we actually get 6 or 8 cores at reasonable prices though - through going this route, AMD has forced Intel to cut into its production margins significantly. Before chiplets, we had 4c4t and 4c8t CPUs in the ~$150-400 range for nearly a decade, while now we get 6C CPUs for less than $200. That's a major value improvement no matter how you look at it.

Of course, per-die costs are a pretty small part of a CPU's price overall - don't they generally hover in the upper double digits range at most?

AusWolfI like small form factor builds. If a CPU runs hot on a test bench, or in a tower case with adequate airflow, then it will definitely throttle in a SFF system.

Not necessarily true. I've seen bench runs of a 5950x running hot, but I've successfully ran one on PBO in a meshilicious. Temps were kept under 80 even with a 6900xt sitting next to it. It depends on the SFF setup. The NZXT 200 with the same setup would overheat quickly, and that's a larger case.

What I'm more interested in is how this story is handled in comparison to Intel chips. Almost all the performance claims you see published are on core performance boost, because the Intel chips need it to be competitive..... but I don't see stories like this about those? (For those at home, I'm being rhetiorical, it's well known that tech journalism has morphed into an extension of marketing and Intel spends quite a bit there). I personally don't care which company I buy, though. I pick the best product for my use.

AerpoweronAfter thinking a while about the posted CBr23 screenshots, we don't even know what the ambient temps were. Cooling solution could the the most crappy Dell stock cooler for all we know, which can barely handle 95W.

But the the 92.1°C temp. I think it must be the single core CBr23 run temp. But it could also be the temp from a Prime 95 single core run, which gets every CPU very hot since you just hit one core with insane loads.
The 110W are likely from the all core Cinebench run. And the temperature at all core load is usually way lower then the single core run temperature. The conclusion is drawn from my Zen 2 and Zen 3 CPUs. Even the R5 3600 behaves in that way.
The 5.45GHz boost could even be the idling boost when CBr23 switches from the multicore test to the single core test. We know these are all the max values, so no reset has been done on HWinfo between the single core and multi core runs.
The score are especially unreliable, since there is variance. HWinfo takes away quite some points from the Single core result, and that is assuming the Windows scheduler didn't throw around the single core load on all cores, which can still happen.
Multi core should be more reliable, but still only one score we see.

@AusWolf I get your concern about the high temperature. But keep in mind, it is most likely under heavy single core load. All core should be much lower. The CPUs are also designed to run at these temps, others already mentioned the heat density you get with these tiny Zen4 cores compared to the relatively large Alder Lake cores.

@phanbuey could you please post the power draw of you 12600K to your Cinebench R23 scores? Not really relevant as comparison, since we don't know anything about cooling, roomtemp or silly stuff running in the background of the 7600X. But i am curious :)

Single :



Multi:

AusWolfThen I really don't understand why users praise chiplets so much, considering that the end product costs the same as traditional, monolithic CPUs. It's only AMD's gain, not ours.

If AMD were doing monolithic designs, to maintain their profit margin they would have to charge far more - or, they'd have to shrink their profit margin and charge the same vs monolithic.

It also allows them to scale up core count greatly, without the big negative impact to yield that the larger monolithic dies have. This is a huge advantage in servers chips in cost, as well as configurability.

But from a pure performance perspective, all else being equal there's no advantage, just the opposite. Look at how long it took them to beat skylake 14nm @ 36MT/mm2 using TSMC N7 @98 MT/mm2.

Now they are on a high power N5 node @ 127MT/mm2 vs Intel 7 92MT/mm2 and based on what I've seen, maybe match up well against Alder Lake in 1T while losing a bit in MT - but not against Rocket Lake.

Intel's real performance issue is not in client, it's really in server where they can't scale up to as many cores as those chiplets can do (Yet).

RandallFlaggIf AMD were doing monolithic designs, to maintain their profit margin they would have to charge far more - or, they'd have to shrink their profit margin and charge the same vs monolithic.

It also allows them to scale up core count greatly, without the big negative impact to yield that the larger monolithic dies have. This is a huge advantage in servers chips in cost, as well as configurability.

But from a pure performance perspective, all else being equal there's no advantage, just the opposite. Look at how long it took them to beat skylake 14nm @ 36MT/mm2 using TSMC N7 @98 MT/mm2.

Now they are on a high power N5 node @ 127MT/mm2 vs Intel 7 92MT/mm2 and based on what I've seen, maybe match up well against Alder Lake in 1T while losing a bit in MT - but not against Rocket Lake.

Intel's real performance issue is not in client, it's really in server where they can't scale up to as many cores as those chiplets can do (Yet).

And that's where Foveros 3D stacking will kick in. Manufacturing those chips is still really hard, though...