TDP is about as meaningful on CPUs now as nm is in fab tech.

In practice the 3950x pulled up to ~225w running maxxed out avx2 workloads, ~300w if you overclocked it. The 3900x pulls up to ~190 watts at stock boosts. Both are called "105w TDP" parts.

PPT on the 105W TDP CPUs are ~143W by default, so your numbers are way of and not possible without overclocking by removing the limit.

You'll also go way beyond 300W with heavy overclocking.

This is incorrect, to the best of my knowledge. 3900X and 3950X both have a 142W package power limit (PPT, 35% above TDP) at stock, with a properly working motherboard. As I understand it, not even short boosts are allowed to use more power than the PPT.

The limit can be lifted in the BIOS (or e.g. Ryzen Master) but that would no longer be stock operation. And there was a defect in the initial BIOS of ASRock X570 and maybe other motherboards, which broke PPT but that was running the CPU out of spec, not stock operation.

If you're going to attack the "spec" that is TDP, you may as well do it correctly. None of the scenarios you mentioned align with what TDP is supposed to "measure". Even if TDP weren't the flawed measure it is, you still wouldn't be critiquing it by noting the measures you have.

Here's the critique you should use! https://www.gamersnexus.net/guides/3525-amd-ryzen-tdp-explai...

Nevertheless, for a given manufacturer, TDP is useful as a coarse grouping of power consumption.

I disagree. If TDP wasn't so flawed as to be completely meaningless, that critique would probably be valid.

The definition AMD uses is absolute nonsense, given how much the θca values differ from model to model. If they fudged any harder the higher end CPUs would start having "TDP" lower than the weak CPUs.

I second this reaction. Historically, clock speeds scaled inversely with the # of cores. Seems like the efficiency is overtaking other constraints at this point.

> Historically, clock speeds scaled inversely with the # of cores.

It's the miracle of chiplets. AMD isn't binning one 16-core CPU, they're binning two 8-core CPUs and then combining them together.

The achievable all-core clock will inevitably scale inversely with # of cores, in practice. The advertised single core and all-core clocks are some combination of binning and pure marketing.

With most modern PC processors, both GPU and CPU, one of the primary limitations is thermal headroom. There are features and technologies with varying names across brands and processors that essentially do the same thing: run at the maximum clock that the current thermal situation will support.

From my personal experience, my Threadripper 3970X will happily maintain ~4.4GHz all-core (rated for 4.5GHz max single-core) so long as I can keep the temperature in or below the 70s, with no overclocking.[0] There are power limits as well, but rated performance numbers are within the power limits. Overclocking can put you past the marked power limits, and certainly needs ample cooling.

[0] Granted, I need to pump cold air into the case to maintain the temperature, but that's a limitation of my current thermal solution. At some point I'll probably upgrade to an excessive water cooling solution (:

That's an interesting piece of data and I'm glad you posted it. I also use a 3970X and can't get it to 4.4GHz even with all but one CCD disabled (much less all-core). I am on air cooling, though, and suspect that switching to water cooling would help a lot. I hit 90C almost instantly under load. (I use the automatic overclocking and can do 4.2GHz all core; much better than the specified 3.7GHz all core.)

Since I don't run into Threadripper owners very often, I'm wondering if yours also has a pretty high idle power? Mine idles at 80-90W (reported; 200+W from the wall) which is surprising to me coming from the Intel world. So much electricity wasted simply because I am too lazy to turn off my computer.

I have a hunch that I won the silicon lottery with it, though I haven't confirmed with any overclocking. I'm happy to dive deeper if you want. Below is a basic summary.

I have 5x140mm intake fans, with a Noctua NH-U14S TR4-SP3 cooler. That runs with push-pull 140mm fans, 2000RPM in/1500RPM out.

This reaches steady state within a few minutes under load. With an open case, it will clock down to ~4.2GHz after 5-10 minutes. With a closed case, that is faster. For fun, I ducted the A/C vent in the room into the case and cranked it. It stayed reliably up around 4.4GHz all core. Technically, I think it would reach a lower steady state clock/higher temp if I left it for days, as it does noticeably warm the room.

I don't recall idle power, but yeah, it's warm.

OK, that's really neat. I have almost the same setup except use 120mm fans on the NH-U14S (to clear the RAM, I populated all 8 slots ;)

I am very tempted to switch to water cooling. Seems like I could make it a little quieter, and get a little bit more performance. $1000 more performance? Probably not. But it would be a fun project.

(In general, I am blown away by the Threadripper's performance. I have C++ builds that took 15 minutes on my 6950X that now take 2!)

My fans are offset ever so slightly to account for RAM clearance, as I, too, have eight sticks. I think we basically built the same computer with a couple different choices (;

I lose millimeters of fan coverage (I have millimeters of fan peeking over the top of the tower), but I opted for this, since I get more airflow over the width of the cooler. The effect is probably measurable in single-digit percentage differences in a highly controlled test.

The big thing for me was the case airflow. That thing throws off so much heat that my case ambient temperature was growing to be 10C or more above the room ambient. Hence the 5x140mm intakes. The back of that machine is a nice heater when I'm crunching anything big.

My goal is to keep those fans, but attached to radiators. I think 5x140mm of radiator should give me much better heat transfer.

I have a 3960X on Gigabyte Aorus Master with an RX480, 1080 Ti, X520 NIC, 3x SSDs, and my total-system idle on Linux is 180–210 W. I do agree that the high idle is frustrating, as it dumps a lot of heat into the room.

That's high! I would expect around 100W without the NIC and the second GPU. But it's high even when accounting for that.

You could undervolt slightly, it usually results in less power consumption and higher performance as a result of lower temperature.

Got a few of these and water absolutely make a huge difference. Open air chassis with a large radiators I intended to cool video cards with... not that one can be found right now. Enclosures were turning mine into an easybake. Surfing at 46C, mid 60s pushing it hard. The dust bunnies can get ugly, however.

https://i.imgur.com/SpMMPXQ.jpg

Yeah, you don't want to hit 90 degrees on Zen. Unlike many other CPUs, max stable frequency on Zen clearly decreases with high temps. Time to build a water loop :)

re: idle, you do have pstates enabled, right? ("Cool'n'Quiet") firmware likes to disable it when you overclock in the setup screen, have to turn it back on

Yup. Hitting 90C is basically what limits the top speed, so I know that better cooling might help.

I do have pstates enabled. Power and frequency scale up as you'd expect, in response to system load.

The thermal wall off 100W/cm² is not much about how much heat you can sink, but how much you produce.

The heat cannot leave the silicon itself quick enough with modern chips.

The difference with watercooling for CPUs seems to mostly come down to lower noise, and only marginally better thermals.

I can get much more surface area of radiator than can reasonably be reached by the heat pipes on an air cooler. There are limits to how many radiators you can effectively leverage in a cooling loop.

Additionally, a loop with a large volume of liquid offers much more thermal buffer before reaching a steady state temperature.

Most air coolers will be heat saturated within a minute or so. A water cooling loop may maintain lower temperature for minutes to tens of minutes. So even with two solutions that otherwise reach the same steady state temperatures (and therefore throttle equally), you may see better real world performance out of the water cooling solution.

I'll note, I would be building an open loop, not using an AIO/closed loop cooler. My case has room for 7x140mm of radiator in a couple of configurations. I would probably use 5x of that in one 420mm radiator and a 280mm radiator. This should offer much more cooling capacity than any tower cooler.

Watercooling has one benefit: The heat capacity* of all those water is amazing.

*Heat capacity = how many joules of energy the whole system takes to raise the system's temperature by 1C.

So, the CPU is generating x Watts of heat = x Joules of heat energy per second. Watercooling can absorb a LOT of Joules before the cooling system's temperature increases, thus helping in keeping the CPU away from its MaxTemp.

Plus when attached to a sizable radiator, watercooling system can release all those energy to the environment faster, thus reducing the effective energy absorbed by the system.

I can only offer anecdata, but I built a Ryzen 3600 based desktop earlier in the year.

Initially I used the stock cooler, but it idled at ~45C, and the moment I did anything approaching a load it immediately shot to 90C. This was in a room with ambient at around 23C.

After getting annoyed for a while I swapped for an AIO Liquid cooler and hey-presto it now idles at 30C and when maxed out - 75C.

That's not really a good air vs water comparison. You'd have gotten similar results if you swapped for a better tower cooler as well. The stock coolers are basically built to cover base clock and a bit of turbo.

Yeah with a CoolerMaster Hyper 212 and a single fan I get idle of 40-45 and a high of 76 and I tend to run with fans set to be mostly silent. With a bit of a noisier experience I seem to recall idle was under 40.

That is all OK, but remember that Zen 2 really tries hard to hit 90C. Especially if you go into Ryzen Master and turn on Automatic Overclocking (very conservative overclocking by enthusiast standards), and your motherboard doesn't have power delivery limitations, the CPU's main goal is to try to hit 90C under load. It will automatically adjust frequency and core voltage up until it hits a thermal limit; by taking more heat away from the CPU with better cooling, you end up with better performance because the CPU will try to automatically give you the best performance that your thermal solution can deliver. AMD didn't document any of this very well (lots of confusion between Precision Boost 2 and Precision Boost Overdrive), but the technology is very impressive.

I guess what I'm saying is that you might be interested in exploring some of these options in Ryzen Master. You have some thermal headroom, so if your motherboard has good power delivery, you can max out the power limits and boost clock limits, and probably squeeze out a little extra performance. Power (and therefore heat) is going to scale quadratically with core voltage, and frequency is going to scale less than linearly with voltage, so you might not see much gain (and might prefer not spinning your fans up to full speed under load). But, it's an option that's available if that sort of thing interests you.

I got about 6% more real-world speed out of my Threadripper 3970X by switching from OC off to Automatic Overclocking; with air cooling. AMD's technology here is pretty impressive.

I didn't see this reply until just now.

I had no idea what Ryzen Master was until you mentioned it - seems it's some AMD utility for Windows.

I'm running a stock hardware configuration, no overclocking, and running Pop! OS 20.04.

Yep. The fewer cores you’ve got, the more heat each can generate (== the more power each can consume), for the same thermal solution.

Maybe not the voodoo you are thinking about but on the GPU side, the 3Dfx Voodoo was indeed a groundbreaking 3D accelerator card.

At the time, 3Dfx main competitors were Nvidia and ATI, Nvidia finally bought 3Dfx and AMD bought ATI. So technically, voodoo is on the side of AMD's competition.

Intel's name for its CPU architecture is "Core." Nevertheless, AMD discusses its CPU cores. "voodoo" and "Voodoo" need not be the same (;

Should have named their CPU Architecture the "Crore" instead ;-)