Today, I would like to talk about power hungry CPUs and “phantom” throttling, featuring 10980XE. Intel’s 10980XE can require more than 500+W to reliably run at overclock speeds surpassing ~4.7GHz. In our testing we found an issue that was more difficult to spot with the 7980 or 9980. At that time we could hit what we saw as the maximum safe daily overclock on the 7980XE with our cooling solution, so we didn’t experience issues with power delivery to the CPUs to obtain the 4.4-4.7GHz they were running at.
With the 10980XE we run in to these problems with all instruction sets when pushing them just 100MHz higher, with a little bit higher Core Voltage. An issue we ran into was that while they would report running at consistent 4800MHz on SSE or 4600MHz on AVX through torture testing for 8+ hours, they were actually throttling in performance. This is something you would not see in a torture test, and may not even notice in production environments if it doesn’t occur often – but you will see it overtime.
We used Cinebench R15 CPU test to show the difference when looking for this throttling. You’ll see after running the test our Core temperatures were much lower when the CPU was throttling, but it did not deviate from 4600MHz on all the cores while running the test.
Cinebench Result 4.6GHz No Hyper-threading
Then we decided to check for the extreme case, where you decide to purposefully give the CPU less power than it requires to run the benchmark. A few years ago, most would expect to see a Blue Screen of Death, or just to have the application crash – but here the CPU just throttles performance and instead of seeing a result like the above, you get a result you’d see if the cores were running at <3GHz.
Cinebench Result 4.6 GHz No Hyper-threading
Two very different scores, this one at half of the previous. Also running at ~10C lower on all cores.
The issue with this is that there are some benchmarks and tests that you can now say pass while running at 4.8, 4.9, or even 5GHz and beyond on the 10980XE. Since it will just power throttle the CPU and “pass” the test. We saw this while testing at 4.8GHz, CPUs were passing torture testing at <85C max. When adjusting for the power throttling they would start crashing cores or reach upwards of 100C.
It should also be noted that there are ways to ignore this throttling or disable it’s communication from the CPU. While this may work in the short term, it can actually cause real damage to the CPU itself if intended for daily use.
These results can be duplicated on Linux OS’s with various tools, usually sysbench is a good go-to for raw core performance. Phoronix Test Suite also combines a massive variety of tools in to one place with various tools for core, cache, and memory performance. Some noteable ones being IPC_Benchmark, Coremark, 7-zip compression, ramspeed SMP, stockfish 9, stress-ng, and sysbench.
The power hungriness of the HEDT CPUs was apparent when X299 initially released and most motherboards had VRM heatsinks that were just not able to support them – and all of them started arriving with their own fan attachments to try and alleviate the issue.
So with power hungry CPUs, be careful with your overclock, as thermal throttling is not the only thing you need to watch for.