The AMD Threadripper 2 CPU Review: The 24-Core 2970WX and 12-Core 2920X Tested
by Ian Cutress on October 29, 2018 9:00 AM ESTPower Consumption and Uncore Update: Every TR2 CPU
As part of the 2990WX and 2950X review, we published data looking at the core, package, and uncore power delivery on the Threadripper 2 processors. A lot of users were skeptical of our data, which we had expected, and on the basis of comments and discussions, we slightly changed our testing and layout to get a better understanding of what is going on.
Our previous testing involved taking per-core power measurements, and full package power measurements, under a CPU load with a set number of threads. In the first instance, our load was common Prime95, which could be set for a fixed number of threads. Unfortunately, as observed, Prime95 is not a great program on a per-thread level – it works best when all cores are maxed. As a result, we have since moved to POV-Ray in order to make for consistent testing. In most circumstances, POV-Ray seems to push the CPU harder than our Prime95 test ever did, drawing more power. As we cannot adjust the number of threads on POV-Ray, we discovered that through the command line, we can set an affinity mask and limit exactly which cores the program can use.
We go into detail about how to set an affinity mask in our 2990WX CPU0 follow-up article. You can read about it at this link.
In this case, we can select which cores get the POV-Ray threads, and the threads cannot jump to other cores. So as long as the CPU has enumerated the cores appropriately, we can control crossing die boundaries very easily.
The way we gather data is still the same – taking the internal register values for power. This is one of the simplest ways to take power, and is actually the way the system itself determines power limits for reducing clock speeds or reacting to other system stimuli. It is a rough estimate, somewhere within 10% of the real value according to power experts, and may not be as accurate as physical testing, however it has the advantage of giving more data, and in a way that is easily accessible without additional hardware or set up time. With the method we use, we accept those risks and the possibility of variance from real values, however we have confidence that as the CPU uses these values for its own internal metrics, they should also be good enough for us as well.
Another change is in our definitions. Previously we separated the package power into ‘Core’ and ‘Uncore’ (or interconnect, such as Infinity Fabric). AMD’s internal registers actually give the values for total package and per-core power, meaning that the difference between the two values is not only the uncore, but also the IO and the DRAM controllers too. Some Intel chips actually partition out the DRAM and Uncore, however AMD does not, so we have to accept that the number we get is an amalgamation of these values. However, based on the test, the IO and DRAM should be a constant value. But for the sake of clarity, instead of labelling the value ‘Uncore’, it is now labelled ‘non-Core’.
For this update, we re-ran every Threadripper 2 processor with the new test to find the data.
Starting at the Top: 32-Core Threadripper 2990WX
The key stats on the 2990WX is the 32 core design, with a 250W TDP. From our POV-Ray testing, we hit that 250W around 20 cores loaded.
It is worth noting that we see a couple of bumps around 10 cores and 18 cores. If we split out the per-core values, we can see that actually the per-core power goes up for these settings.
The easiest difference to see is at the 16C level, where each core gets on average 9W and then goes up to 10W, despite more load being added. This is because we are crossing die boundaries, and as new threads are being enabled further away from where the others are, it allows more power to be pushed through without disrupting temperatures too much.
The 24-Core Setup: Threadripper 2970WX
Moving down to 24-cores, each of the four silicon dies now has six cores rather than eight, which means we cross those die boundaries earlier. This has benefits and issues worth mentioning.
Now this graph is not as linear as I was expecting. We see big rises at 7-8 core load, 13-14 core load, and 19-20 core load. The reason is the same as the 2990WX: we are crossing die boundaries when more threads are added.
In the per-core charts, this is much easier to see. At the first boundary at 7-8 core load, we go from 12W per core to 15W per core. At the second boundary at 13-14 core load, we go from 10W per core to 12-13W per core. The downside of this is that going past this initial two-core high actually causes a drop in power consumption in order to counteract placing more threads on the same die.
In both the 2970WX and 2990WX, the ‘non-core’ power was around 53-55W of the total package power, even at low loads. This is roughly 20-22% of the total package power. Keep that number for later.
The Previous Recommendation: 16-Core Threadripper 2950X
With only one silicon boundary, we should expect the 2950X to be a reasonably sedate affair. This chip has a maximum TDP of 180W, and fewer IF links to draw power from the total package.
The 2950X power curve under POV-Ray is relatively smooth, and we top out at that 180W TDP value at around 10 cores loaded.
There are no surprises on the per-core values, except that in 1-2 core mode we actually push 21W through those Zen+ cores, compared to 6W per core when the 2990WX is fully loaded. That’s a large range in which the core has to be both efficient and performant.
The 12-Core Backup: Threadripper 2920X
A few years ago, offering a 12-core CPU on a consumer platform would have been a dream come true. It seems odd these days that the lowest CPU in this test is a 12-core. But here we are.
Similar to the 2950X, the power curve is smooth and regular, and we again hit the 180W TDP around 9-10 core load. This time we really need to use all the chip to get to the top power levels.
What is interesting about the per-core values for the 2920X is that it isn’t until 8C fully loaded that we actually see a proper per-core power drop – at one core load, that core is drawing 18W, and at 7C load, the cores are still drawing 17-18W each. That’s a good environment for keeping frequencies high, as I’m sure we’ll see in some of the benchmarks.
Going back to the non-core power, on these last two chips, we see the non-core power around 30W, which is only 16-17% of the total CPU power when fully loaded. This is just over half the power of the quad-die processors, mainly because of fewer Infinity Fabric links.
Peak Power Comparison
So for our peak power comparison tests, we take the peak full package power value from our per-core testing. Again we are using POV-Ray as our load generator.
All the TR2 processors stay within TDP in our testing, but this is down to how AMD defines its TDP.
69 Comments
View All Comments
Ian Cutress - Monday, October 29, 2018 - link
EPYC 7601 is 2.2 GHz base, 3.2 GHz Turbo, at 180W, fighting against 4.2+ GHz Turbo parts at 250W. Also the memory we have to use is server ECC memory, which has worse latencies than consumer memory. I've got a few EPYC chips in, and will be testing them in due course.mapesdhs - Monday, October 29, 2018 - link
Does the server memory for EPYC run at lower clocks aswell?GreenReaper - Wednesday, October 31, 2018 - link
ECC RAM typically runs slower, yes. It's correctness that you're looking for first and foremost, and high speeds are harder to guarantee against glitches, particularly if you're trying to calculate or transfer or compare parity at the same time.iwod - Monday, October 29, 2018 - link
Waiting for Zen2Boxie - Monday, October 29, 2018 - link
only Zen2? Psshh - it was announced ages ago... /me is waiting ofr Zen5 :Pwolfemane - Monday, October 29, 2018 - link
*nods in agreement* me to, I hear good things about Zen5. Going to be epyc!5080 - Monday, October 29, 2018 - link
Why are there so many game tests with Threadripper? It should be clear by now that this CPU is not for gamers. I would rather see more tests with other professional software such as Autoform, Catia and other demanding apps.DanNeely - Monday, October 29, 2018 - link
The CPU Suite is a standard set of tests for all chips Ian tests from a lowly atom, all the way up to top end Xeon/Epyc chips; not something bespoke for each article which would limit the ability to compare results from one to the next. The limited number of "pro level" applications tested is addressed in the article at the bottom of page 4."A side note on software packages: we have had requests for tests on software such as ANSYS, or other professional grade software. The downside of testing this software is licensing and scale. Most of these companies do not particularly care about us running tests, and state it’s not part of their goals. Others, like Agisoft, are more than willing to help. If you are involved in these software packages, the best way to see us benchmark them is to reach out. We have special versions of software for some of our tests, and if we can get something that works, and relevant to the audience, then we shouldn’t have too much difficulty adding it to the suite."
TL;DR: The vendors of the software aren't interested in helping people use their stuff for benchmarks.
Ninhalem - Monday, October 29, 2018 - link
ANSYS is terrible from a licensing standpoint even though their software is very nice for FEA. COMSOL could be a much better alternative for high-end computational software. I have found the COMSOL representatives to be much more agreeable to product testing and the support lines are much better, both in responsiveness and content help.mapesdhs - Monday, October 29, 2018 - link
Indeed, ANSYS is expensive, and it's also rather unique in that it cares far more about memory capacity (and hence I expect bandwidth) than cores/frequency. Before x86 found its legs, an SGI/ANSYS user told me his ideal machine would be one good CPU and 1TB RAM, and that was almost 20 years ago.