By: -.- (blarg.delete@this.mailinator.com), August 24, 2018 8:08 pm
Room: Moderated Discussions
Travis (travis.downs.delete@this.gmail.com) on August 24, 2018 4:49 pm wrote:
> You can find some generously contributed Gold 6130 results linked from this github issue.
>
> This of course is a more interesting machine than an obscure W-2104.
>
> The same pattern of behavior is observed: even "heavy" AVX-256/AVX-512 instructions
> don't get the heavy downclock, unless a lot of them are issued in parallel.
>
> What is weird though is that the published non-AVX/AVX/AVX-512 turbos for that box are
> 3.7/3.6/3.5 GHz (for 1 or 2 cores), but the test fairly clearly shows that the achieved
> speeds are in fact 3.7/3.5/3.3, so a 200 MHz delta between each tier not 100 MHz.
>
> So maybe you aren't getting what you thought you were paying
> for on the Gold 6130? Or something else is going on...
The link to the results is dead, but how much of a role does the motherboard play? On Core-X at least, the AVX frequency offsets are configurable via BIOS. I imagine that this is more restricted on Xeon, but presumably the motherboard still plays a role there.
Trying your tool, one thing I notice is that the the AVX frequencies aren't different on 'favored' cores. For this 7820X running at stock frequencies, running 1 thread on a non-favored core, I get 4.3/3.8/3.6 GHz, whilst on a favored core, I get 4.5/3.8/3.6 GHz. Note that I'm running this using the --no-pin option, and pinning via taskset instead. I do get some weirdness when trying more than one thread being pinned on different cores.
I don't think any of the Xeons have Turbo Boost 3.0, so this distinction is probably just a Core-X thing.
> You can find some generously contributed Gold 6130 results linked from this github issue.
>
> This of course is a more interesting machine than an obscure W-2104.
>
> The same pattern of behavior is observed: even "heavy" AVX-256/AVX-512 instructions
> don't get the heavy downclock, unless a lot of them are issued in parallel.
>
> What is weird though is that the published non-AVX/AVX/AVX-512 turbos for that box are
> 3.7/3.6/3.5 GHz (for 1 or 2 cores), but the test fairly clearly shows that the achieved
> speeds are in fact 3.7/3.5/3.3, so a 200 MHz delta between each tier not 100 MHz.
>
> So maybe you aren't getting what you thought you were paying
> for on the Gold 6130? Or something else is going on...
The link to the results is dead, but how much of a role does the motherboard play? On Core-X at least, the AVX frequency offsets are configurable via BIOS. I imagine that this is more restricted on Xeon, but presumably the motherboard still plays a role there.
Trying your tool, one thing I notice is that the the AVX frequencies aren't different on 'favored' cores. For this 7820X running at stock frequencies, running 1 thread on a non-favored core, I get 4.3/3.8/3.6 GHz, whilst on a favored core, I get 4.5/3.8/3.6 GHz. Note that I'm running this using the --no-pin option, and pinning via taskset instead. I do get some weirdness when trying more than one thread being pinned on different cores.
I don't think any of the Xeons have Turbo Boost 3.0, so this distinction is probably just a Core-X thing.
