90c+ CPUs

[Zach_01]

Transistors and size

Zen3 8-core CCD: 4.15B, 83mm²
Zen4 8-core CCD: 6.57B, 70mm²

58.3% more transistor in a 15.3% smaller area.

Zen3: 50M/1mm²
Zen4: ~94M/1mm²

Almost double (+88%) transistor density

Even at the exact same power heat dissipation is getting harder indeed.
Let alone the increased power limit
142W >> 220~230W

The 115C Tjmax is not so crazy now I guess. I mean… it is but, no wonder…

 

[Deleted member 185158]

Transistors and size

Zen3 8-core CCD: 4.15B, 83mm2
Zen4 8-core CCD: 6.57B, 70mm2

58.3% more transistor in a 15.3% smaller area.

Zen3: 50M/1mm2
Zen4: ~94M/1mm2

Almost double (+88%) transistor density

Even at the exact same power heat dissipation is getting harder indeed.
Let alone the increased power limit
142W >> 220~230W

The 115C Tjmax is not so crazy now I guess. I mean… it is but, no wonder…

105w and 6.57B transistors, how many watts does a single transistor require to operate?

 

[Zach_01]

105w and 6.57B transistors, how many watts does a single transistor require to operate?

I think it’s impossible to calculate this.

Between different core speeds and loads, different characteristics of cores, caches and what ever else is in there, dynamic operation with possibly different rate for each section of the CCD (caches, cores, etc).
Also traces in between them could add some more to resistance…

…and I’m sure I can’t think of everything.

Any attempt of calculation would terribly wrong.

Of course you can find an average value if 105/6.57B that includes every resistance possible inside the 8-Core CCD but you can’t really know what can be the min/max value for a singe transistor at any given time even for a locked speed and power.

I mean even at static conditions (speed, voltage, power) you can’t really know how the CPU distributes anything in there and what parts are (and if) working more than others.

 

[Deleted member 185158]

I think it’s impossible to calculate this.

Between different core speeds and loads, different characteristics of cores, caches and what ever else is in there, dynamic operation with possibly different rate for each section of the CCD (caches, cores, etc).
Also traces in between them could add some more to resistance…

…and I’m sure I can’t think of everything.

Any attempt of calculation would terribly wrong.

Of course you can find an average value if 105/6.57B that includes every resistance possible inside the 8-Core CCD but you can’t really know what can be the min/max value for a singe transistor at any given time even for a locked speed and power.

I mean even at static conditions (speed, voltage, power) you can’t really know how the CPU distributes anything in there and what parts are (and if) working more than others.

Yea, you're right. Probably like 0.0000000000000000000000000000000000000000000000000001234w or some shit.

At least what we have today beats what was going on back in school the early 90s.
Like 10 mhz chips. Blazing fast computing. Maybe 5w and 100K transistors maybe. Everything off die basically. Cache upgradeable expandable.

But luckily, you could break down some of that wattage usage from HWInfo64.

The 12400F breaks down

_________________________Max
CPU package power W - 62.579
IA cores Power W - 57.823
SA power W - 3.383
Rest of chip power W 0.449

Completing WPrime 32m in 2.7s (no avx)
DDr5 memory sticks report 1w and 0.875w with low of 0.125w. 2759mhz
Heat sinks on memory sticks? I don't think it's needed here....

Makes me curious at what frequency and or core reduction these new 7000 series chips become able for passive cooling.

 

[Count von Schwalbe]

I wonder if we will begin to see vapor chamber IHS in future AM5 CPUs. The heat density is considerable, and the very thick IHS could make it a possibility.

 

[Night]

Can you see how that doesn't make sense at all. Maybe processor performance but not watt power. 100w is 100w.

Yes, however 5950X for example boost clocks to 4.9 GHz out of the box, while 7950X boosts to 5.7 GHz, that's a substantial increase for the same amount of power being used, a 7950X would use around 80W or less for a clock of 4.9 GHz. As far as cooling goes, a physically bigger chip and its IHS would be more efficient at transfering heat in my opinion, making it easier on cooling solutions.

 

[AusWolf]

As far as cooling goes, a physically bigger chip and its IHS would be more efficient at transfering heat in my opinion, making it easier on cooling solutions.

That! When you increase your cooler's size, you're increasing its contact surface with air. When you increase the size of a chip, you're increasing its contact surface with the IHS / cooler coldplate.

 

[tabascosauz]

The 12400F breaks down

_________________________Max
CPU package power W - 62.579
IA cores Power W - 57.823
SA power W - 3.383
Rest of chip power W 0.449

Completing WPrime 32m in 2.7s (no avx)
DDr5 memory sticks report 1w and 0.875w with low of 0.125w. 2759mhz
Heat sinks on memory sticks? I don't think it's needed here....

Makes me curious at what frequency and or core reduction these new 7000 series chips become able for passive cooling.

I used to have this whitepaper saved somewhere about how Intel hardware internally measures power. Off the top of my head I remember Intel measures not estimates most rails.

On AMD only Vcore and VSOC are measured, everything else is estimated to add up to Package power. Which leads to some wacky numbers that don't add up on APU, but in the end measured EPS power is still roughly in the right place.

Both can have their power reporting messed with, though.

Manufacturers will always say measure EPS if you want accurate numbers but that includes VRM losses, not sure how much difference it makes. But, it should be very close (1-2W tops) if my 2060 Super is any indication - the total board input power is usually about a watt off the GPU power (relatively mediocre 8(?)phase VRM with ONsemi 50A parts). It's a shame that the total input power measurement disappeared on RTX 30 series, but AFAIK GPU power is still measured the same way (albeit transients are now even worse as with RDNA2).

 

[AusWolf]

Here we go.

 

[Zach_01]

@buildzoid has a video with testing 7950X power limits with a little unusual method. He compares the CPU power scaling against it self on 50W as a starting point up to 275W (steps of 25W).
Its interesting, as it shows that past 175~200W there is almost no gains.
I think most of us here in this thread at least already established that.

A reminder that this test is with just power limit increase and nothing else (no info about EDC either). Its not that the CPU necessarily did draw that level of power (275W).
In fact there is no info of what level of power the CPU did reach but given the results I would say around 200~210W.

Remember also that Ryzen7000 is very much.... I want to say temp sensitive, but not quite... I'll say heat dissipation sensitive first and then temp sensitive.
Per AMD: The CPU will reach temp limit before power limit.
Most likely if there is a possibility for a higher dissipation rate (better cooler) the CPU would scale a little better past the 150~175W points and up to stock power limit (220~230W). Not great, sure... but better.

 

[AusWolf]

@buildzoid has a video with testing 7950X power limits with a little unusual method. He compares the CPU power scaling against it self on 50W as a starting point up to 275W (steps of 25W).
Its interesting, as it shows that past 175~200W there is almost no gains.
I think most of us here in this thread at least already established that.

A reminder that this test is with just power limit increase and nothing else (no info about EDC either). Its not that the CPU necessarily did draw that level of power (275W).
In fact there is no info of what level of power the CPU did reach but given the results I would say around 200~210W.

Remember also that Ryzen7000 is very much.... I want to say temp sensitive, but not quite... I'll say heat dissipation sensitive first and then temp sensitive.
Per AMD: The CPU will reach temp limit before power limit.
Most likely if there is a possibility for a higher dissipation rate (better cooler) the CPU would scale a little better past the 150~175W points and up to stock power limit (220~230W). Not great, sure... but better.

"Not great, not terrible." 95 °C is the new 3.6 Roentgen.

 

[Deleted member 185158]

I used to have this whitepaper saved somewhere about how Intel hardware internally measures power. Off the top of my head I remember Intel measures not estimates most rails.

On AMD only Vcore and VSOC are measured, everything else is estimated to add up to Package power. Which leads to some wacky numbers that don't add up on APU, but in the end measured EPS power is still roughly in the right place.

Both can have their power reporting messed with, though.

Manufacturers will always say measure EPS if you want accurate numbers but that includes VRM losses, not sure how much difference it makes. But, it should be very close (1-2W tops) if my 2060 Super is any indication - the total board input power is usually about a watt off the GPU power (relatively mediocre 8(?)phase VRM with ONsemi 50A parts). It's a shame that the total input power measurement disappeared on RTX 30 series, but AFAIK GPU power is still measured the same way (albeit transients are now even worse as with RDNA2).

For earlier gen Ryzen chips, the HWInfo power deviation reporting at or near enough to 100% should have the PPT reading pretty close to accurate. give or take 5% I believe, which is also the allowable (in most processors) TDP swing. This would be in regards to total package power draw.

If the chip is 145w power draw limit, depending on how the bios is written, this figure can be skewed however. Basically the bios lying and reporting a lower figure than actual, the power deviation would not be near 100%, then you know something whacky is going on.

@buildzoid has a video with testing 7950X power limits with a little unusual method. He compares the CPU power scaling against it self on 50W as a starting point up to 275W (steps of 25W).
Its interesting, as it shows that past 175~200W there is almost no gains.
I think most of us here in this thread at least already established that.

A reminder that this test is with just power limit increase and nothing else (no info about EDC either). Its not that the CPU necessarily did draw that level of power (275W).
In fact there is no info of what level of power the CPU did reach but given the results I would say around 200~210W.

Remember also that Ryzen7000 is very much.... I want to say temp sensitive, but not quite... I'll say heat dissipation sensitive first and then temp sensitive.
Per AMD: The CPU will reach temp limit before power limit.
Most likely if there is a possibility for a higher dissipation rate (better cooler) the CPU would scale a little better past the 150~175W points and up to stock power limit (220~230W). Not great, sure... but better.

That was really interesting.

The high temperature threshold specifies the CPU temperature that causes ALERT_L to assert if the CPU temperature is greater than or equal to the threshold. SBTSI::HiTempInt and SBTSI::HiTempDec combine to specify the high temperature threshold. See 6.2.5 [SB-TSI Temperature and Threshold Encodings]. Reset value equals 70 °C. Write access causes a reset of the alert history counters (specified by SBTSI::AlertThreshold[AlertThr]) and the corresponding timer (specified by SBTSI::UpdateRate[UpRate]). See 6.2.3 [Alert Behavior].

That would be page 322 on the white sheet I left a couple pages back.

This specific quote is what I look for in the white papers. The above is the same in all generations of Ryzen chips available, but I cannot confirm this for Zen 7000.

I would pay close attention to the threads and comments, "My CPU sounds like a jet engine!!!"" and the above is why.

100% temperature sensitive. The cpu would use the temperature change given over a small threshold of time and will adjust power and frequency accordingly.

Probably with some other values involved, but 70c, the Cpu will not request additional power.

_____

So hopefully at some point they release the white papers.

The revision guide is viewable however.

 

[SL2]

I wonder if we will begin to see vapor chamber IHS in future AM5 CPUs. The heat density is considerable, and the very thick IHS could make it a possibility.

Or, what if that beefy lid is a placeholder for a stack of multiple dies in the future, not just CCD + 3D cache. I've heard about such designs for years, but never understood how they would cool it properly..

Maybe the lid will get thinner even after 7000X3D and before AM5 is EOL, because of more stacked dies?

If AMD just wanted to keep the height without making a thick lid, why didn't they? They designed the socket from ground up, I fail to see what could have stopped them. AM4 coolers are nowhere near 100 % compatible anyway, so that's not it.

On the other hand, AMD must make the lid thicker now in order to make room for stacked dies in the future, without having to change the socket specifications later.
(The CCD in 5800X3D is thinner than other CCD's because otherwise there would not be any room for the cache die.)

Maybe the thick lid is just a sign of something yet to come, and not just an annoying compromise, or a mistake.

 

[A Computer Guy]

@buildzoid has a video with testing 7950X power limits with a little unusual method. He compares the CPU power scaling against it self on 50W as a starting point up to 275W (steps of 25W).
Its interesting, as it shows that past 175~200W there is almost no gains.

I'd like to see that same test for comparison with a delidded Zen4 CPU.

Maybe the thick lid is just a sign of something yet to come, and not just an annoying compromise, or a mistake.

Maybe something juicy like stacked e-cores. That would be wild.

 

[Deleted member 185088]

@buildzoid has a video with testing 7950X power limits with a little unusual method. He compares the CPU power scaling against it self on 50W as a starting point up to 275W (steps of 25W).
Its interesting, as it shows that past 175~200W there is almost no gains.
I think most of us here in this thread at least already established that.

A reminder that this test is with just power limit increase and nothing else (no info about EDC either). Its not that the CPU necessarily did draw that level of power (275W).
In fact there is no info of what level of power the CPU did reach but given the results I would say around 200~210W.

Remember also that Ryzen7000 is very much.... I want to say temp sensitive, but not quite... I'll say heat dissipation sensitive first and then temp sensitive.
Per AMD: The CPU will reach temp limit before power limit.
Most likely if there is a possibility for a higher dissipation rate (better cooler) the CPU would scale a little better past the 150~175W points and up to stock power limit (220~230W). Not great, sure... but better.

Seems AMD went full on to take on Intel, with Zen3 it seems they left some performance on the table, have they been more forgiving power and temperature wise they could've been faster.
Having said so, I do think we have this unfortunate trend of pushing hardware beyond its sweet spot, especially with some "reviewers" giving too much importance for 2 more FPS as if there are no settings to tweak (useless RT alone gives 50%+ performance).