Hey guys,
So as the title says, I am using two identical systems, with identical drivers and os. Both have ryzen threadripper 5965x. When I run geekbench 5, I get 18.5k score on multi core with one system and with other I get around 14k. I have checked for temperatures, and I do not recognize thermal throttling…I mean under complete load I observed 79 degrees tops. Also, I experimented with scheduling governor and set it to ‘performance’ setting, but it did not improve anything. Under full load, all threads work on 4.4ghz.
–single core performance is identical for both.
What could be the cause for such a large gap in performance? I can paste the link to geekbench score.
Hello, thank you for your answer. Both systems are being run without display, so the access is via ssh. Both have exact same ram module types and same storages. As I said, both have identical configs.
Since Jode pointed out all those things I’m gonna add a new element to the puzzle: are you running the same motherboard with the same BIOS version on both systems?
Ryzen CPUs tend to boost more dynamically like GPUs so if one system is running at 79°C while the other is running at, for example, 60°C you could be loosing some performance. Though the performance discrepancy should still not be that big. Have you checked that the CPU is topping out at 79°C instead of throttling down and stabilizing to 79°C?
Thank your for your answer. Both motheboards are same, as mentioned earlier;both configs are identical. BIOS versions are same, as well. If you need guys, I can add some logs.
If both pc’s are running at 79 deg but one is performing better than it still could be a cooling issue. Can you look at the clock speeds using ryzen master? That will tell you what the cpu is actually running at, you can also lock the clockspeed to the same and then see what the temperatures do. If one is significantly cooler then that would be the difference, the other way is that one has lower voltage which would mean one cpu is a better chip.
I think you are running linux, can you log anything like temperatures, clocks and voltages over the course of a run?
I’m seeing mostly a large multi core difference with especially the crypto score being about 4 times as high for the other. where others are close or the diference not as large.
Looking at the aes-xts test, i think it uses avx512, which is also very power hungry. But it’s not very memory bandwidth sensitive. Single core is exactly the same so avx512 is probably enabled. Can you check power use during the test? I think the power limit might be different between the 2 computers. (or maybe pbo)
I have experimented with different scheduling, i.e. I used performance mode, to try that out. However, now when I run sensors command I am getting N/A for Adapter:virtual device temperature on problematic system, and around 40 degrees on the the other one. Also, I ran the test again and will provide the temperatures along side the scores now.
Power limit isn’t set by performance mode, it’s a setting in the bios. There’s also a temperature limit, which is probably around 80c or so.
I’m not sure if there’s software to control AMD CPU power limits. AMD’s windows software might, but it sounds like you’re using linux?
You may want to also make sure to use the amd_pstate frequency governors for more optimal performance, though you would have to validate performance in your workload vs stock acpi_cpufreq. Active or guided are recommended over passive for best results, I believe.
