AMD Ryzen 9 3950X

[scaryjim]

But also.. not like for like. There has been amazing progress, but the TDP 'rating' has vastly changed as well. Intel's 95W is actually 210W TDP (both intel figures) for PL2 state ...

I sort of take your point, but it's also worth noting that the PL1 95W figure is (theoretically, at least) for all cores @ 3.6GHz - so it's surprisingly close to like-for-like in that regard. And iirc AMD's TDP figures are the power targets for maximum all-core turbo, so they're much more comparable to old TDPs @ set stock frequency

[DanceswithUnix]

If you think that those 115W TDPs were for single core hyperthreaded processors running at around 3.6GHz, it throws a lot of light on just how much progress has been made in the last 15 years...!

Some of that improvement was from going *back* to Pentium III as the underlying architecture for the Core series. P4 was just bad.

I sort of take your point, but it's also worth noting that the PL1 95W figure is (theoretically, at least) for all cores @ 3.6GHz - so it's surprisingly close to like-for-like in that regard. And iirc AMD's TDP figures are the power targets for maximum all-core turbo, so they're much more comparable to old TDPs @ set stock frequency

AMD's TDP figures are confusingly done. My 3700X can pull 90W at stock on the EPS12V rails running blender despite the 65W TDP. That's a few watts short of the 9900K. (from https://www.gamersnexus.net/hwreview...marks-vs-intel)

So yeah, gaming etc the CPU stays within TDP, but AVX workloads soon crank it past.

[watercooled]

With regard to those moving window power targets, it does seem a bit like it's a convenient way of pretending TDP is lower when they know full well that enthusiast motherboard manufacturers will set the time to infinity.

On the flip side, I've owned a few Intel CPUs where the real-world power consumption never even gets close to the TDP. E.g. a Sandy Bridge Celeron (can't remember whether G530 or G540) in a small server I set up has a 65W TDP, and I can't remember where I recorded the measured power consumption but IIRC it was something like 40W (maybe not but definitely it was definitely lower than the TDP) at the wall under power virus load.

TDP has been an overly general approximation of power consumption for as long as I can remember on desktop CPUs, it's just become more pronounced recently when real-world power consumption is miles higher than TDP would suggest, leading to confusion with powering and cooling solutions for end users.

You also get arguments about heat produced isn't the same as power, which is totally untrue, or at least a dangerous over-simplification. Given CPUs haven't started breaking the laws of thermodynamics, power in = heat out, minus maybe a tiny bit in sound or electromagnetism if you're being pedantic. It kind of makes sense when you talk about averages over time due to thermal masses, but simply claiming heat is not power is just wrong.

[scaryjim]

... TDP has been an overly general approximation of power consumption for as long as I can remember on desktop CPUs, it's just become more pronounced recently when real-world power consumption is miles higher than TDP would suggest, leading to confusion with powering and cooling solutions for end users. ...

That kind of makes sense though.

Historically I doubt TDPs were ever meant for end users - I'm pretty sure they'll have been for OEMs to help them specify cooling solutions in pre-built PCs. And up to around 2004ish both Intel and AMD specified precise TDPs for each chip spec, often to within 0.1W.

Then somewhere around 2004 they started using grouped TDPs, where several similar processors were all assigned a single TDP. And I don't believe that those different-specced chips all magically started producing the same, nice, round wattages - it made it easier for OEMs to design cooling solutions.

And now, the on-chip monitoring and management is sufficiently advanced to adapt to the cooling available anyway. The TDP values are still relevant to OEMs who want to know whether they can stick a given CPU in a given chassis (my SFF case can dissipate 95W? Great, I can slap a 9900k in it.), and to consumers who need to know how big a cooler they should buy. And it's not like they advertise a base frequency that can't be acheived with TDP-spec cooling.

As to power draw? Well, TDP isn't a measure of power drawn. It never has been. It was never intended to be. And even thought it's a VERY common misconception, that's hardly Intel/AMD's fault. So I'm not going to jump on their backs for that one.

[kalniel]

As to power draw? Well, TDP isn't a measure of power drawn. It never has been. It was never intended to be. And even thought it's a VERY common misconception, that's hardly Intel/AMD's fault. So I'm not going to jump on their backs for that one.

Certainly not any kind of measure of expected power draw. But peak/time window power draw? I'll bite: what's the difference between the thermals a system has to dissipate and power draw of said system?

[DanceswithUnix]

Certainly not any kind of measure of expected power draw. But peak/time window power draw? I'll bite: what's the difference between the thermals a system has to dissipate and power draw of said system?

If you make a device that spends half of its time at 200W and half of its time at 0W then it requires a 200W PSU, but requires a 100W thermal solution. I think that is where some of the gripes come from. On top of that, the TDP numbers are really just made up and according to the heatsink vendors aren't actually useful for choosing a heatsink, which has to make you wonder what good it really is.

[watercooled]

And now, the on-chip monitoring and management is sufficiently advanced to adapt to the cooling available anyway. The TDP values are still relevant to OEMs who want to know whether they can stick a given CPU in a given chassis (my SFF case can dissipate 95W? Great, I can slap a 9900k in it.), and to consumers who need to know how big a cooler they should buy. And it's not like they advertise a base frequency that can't be acheived with TDP-spec cooling.

That's debatable though, especially when using actual, real-world motherboards which pretty much ignore those specifications and would have the CPU constantly bouncing off throttling points if you only allowed for that cooling capacity. As for knowing whether a whole system is capable of dissipating 95W of heat and maintaining a given air temperature supplied to the cooler? Yeah, good luck with that. It's at the point where it's a near-useless, arbitrarily-calculated number stuck on packaging for product segmentation/marketing. There's a reason CPUs with 'equal' TDPs are often supplied with different coolers by the manufacturer - even they know it's a load of tosh!

As to power draw? Well, TDP isn't a measure of power drawn. It never has been. It was never intended to be. And even thought it's a VERY common misconception, that's hardly Intel/AMD's fault. So I'm not going to jump on their backs for that one.

I worded that part of my post very carefully. But it's as much a measure of power drawn as it is a measure of heat produced given they are the same thing - CPUs do not defy the laws of thermodynamics. It may be a measure of average thermal dissipation required over a moving window, but by the same logic that must mean average power over that same window. It's not really a misconception when it's a simple and irrefutable fact. Power drawn (in Watts if you want to give it a unit), is equal to power dissipated (also in Watts - it's a measure of power, not exclusively electrical power) at any given moment.

Extrapolating from that information may lead to incorrect assumptions, which is arguably not helped by the manufacturers given the constantly changing and totally arbitrary definitions. Reviews judging efficiency on that number also helps to fuel the confusion.

[fend_oblivion]

AMD is doing extremely well on the CPU front. Now all they have to do is up the ante on the GPU side.

[kalniel]

If you make a device that spends half of its time at 200W and half of its time at 0W then it requires a 200W PSU, but requires a 100W thermal solution.

In that example TDP *is* a measure of power draw - you've drawn 100W/time and that's the same as the thermal solution you need. But yes, I get the point about instantaneous draw might be over 100W - but likewise you might need to extract more than 100W of heat too - albeit you probably have a buffer.

I think that is where some of the gripes come from. On top of that, the TDP numbers are really just made up and according to the heatsink vendors aren't actually useful for choosing a heatsink, which has to make you wonder what good it really is.

Agreed - though I do like AMDs stepdown option as an easy way to run a system cooler/quieter.

[scaryjim]

That's debatable though, especially when using actual, real-world motherboards which pretty much ignore those specifications and would have the CPU constantly bouncing off throttling points if you only allowed for that cooling capacity. ...

This was kind of the point I was making though. They advertise the base frequency as the one you'll end up hitting under baseline thermal constraints. That 3.6GHz is the speed you should get when the system 'throttles' for thermals - i.e. it's not really throttling, it's just not boosting past its base frequency.

Now, if the thermal throttling drops the system below that base frequency, that's more of an issue.

[DanceswithUnix]

In that example TDP *is* a measure of power draw - you've drawn 100W/time and that's the same as the thermal solution you need.

But a 100W psu would go bang. The thermal inertia of a reasonable sized lump of copper can soak up power from a cpu on a timescale measured in seconds, so you can take some liberties with it. The input voltage is very tightly regulated though, you have to give the CPU whatever input power it wants, and you have to keep up to a timescale measured in microsoconds.

Still, that was a rubbish and hurriedly conjured exampled, I'll try something more real world...

You have a CPU that during gaming fluctuates but averages about 60W, during heavy SSE/AVX workloads is pulls 87W, and you have a warehouse full of suitable 125W coolers so you bundle one of those. So what is the TDP??

If I was talking about an Athlon 1000, then in those days max power consumed was the TDP so it would be an 87W part. But a different TDP for every part was considered too complicated for us dumb users.
If I was talking about an Athlon 64 where the chips were rated by which of the standard coolers were deemed necessary, it would be a 125W part. But calling it a 125W part would get you crucified in reviews.
So AMD round up the power required gaming to 65W, and that is what they call the 3700X.
Is it a useful figure? That's dubious. Is it the amount of heat you need to get rid of or the amount of power you need to feed it? No.

Of the possible numbers to choose: 60, 87, 125; they chose 65, another number entirely. They made it up. Intel are just as bad, if anything slightly worse.

[watercooled]

This was kind of the point I was making though. They advertise the base frequency as the one you'll end up hitting under baseline thermal constraints. That 3.6GHz is the speed you should get when the system 'throttles' for thermals - i.e. it's not really throttling, it's just not boosting past its base frequency.

Now, if the thermal throttling drops the system below that base frequency, that's more of an issue.

Yeah I get where you're coming from, and I could be wrong but I think the way boosting works i.e. the CPU running at a higher voltage and less efficiently, it can end up at lower than base clocks under thermally-limited conditions where the motherboard is still allowing unrestricted power draw.

That's something we see on phones too. E.g. where you have two phones with the same SoC, one has a more restrictive power profiles which actually allows higher sustained clocks vs the unrestricted power profile where the CPU is constantly oscillating around the throttle point.

Or in another example, the same phone with different schedulers: https://www.anandtech.com/show/12620...xy-s9-part-2/3

[Ravuncher]

Quality review, the demand is too damn high. AMD aimed to please...

[the_Unforgiven]

May I have one for Christmas please