Does anyone know if Skylake-X CPUs are suffering from the same AVX offset bug that affects Coffee Lake? I wonder if this is intentional behavior or Intel will ever fix it with a microcode update?
AVX Offset Bug
chris719
1 - 20 of 21 Posts
Joined
·
11,217 Posts
I don't think it's a bug. Essentially it downclocks to the AVX offset speed when a supposed non-AVX program runs.
https://forums.overclockers.co.uk/threads/8packs-no-bs-testing-intel-8700k-and-z370-with-delid-5ghz-and-above-available-from-ocuk.18795256/page-20#post-31258558
"CPU runs AVX clocks if AVX offset is set, despite no AVX instructions being used to our knowledge. Seems to happen happen a lot in games. "
https://forums.overclockers.co.uk/threads/8packs-no-bs-testing-intel-8700k-and-z370-with-delid-5ghz-and-above-available-from-ocuk.18795256/page-20#post-31258558
"CPU runs AVX clocks if AVX offset is set, despite no AVX instructions being used to our knowledge. Seems to happen happen a lot in games. "
Some very newly compiled applications may contain AVX instructions, but most games really aren't running any tight loops with vector instructions. Seems like it should switch to the AVX offset multiplier once AVX instructions are fetched from the instruction cache and bring itself out of the AVX offset multiplier as soon as the AVX instructions are done executing. Pretty sure that's how it worked on my Broadwell-E. I don't recall seeing it ever activate AVX offset except when stress testing with P95.
Quote:
This means using just a handful of AVX 256-bit instructions can result in lower performance.
To check if the AVX offset is happening because of a bug or because there really are 256-bit AVX instructions then maybe one could try to see what happens if the Windows "BCDEdit /set xsavedisable 1" command is used but be aware if the software is not properly checking for OS AVX support there can be some crashes.
An example of a user wanting to try disabling AVX for other reasons, namely the bclk exploit in Skylake of which one of the effects is to never power on the 256-bit registers.
https://forums.geforce.com/default/topic/949301/how-to-disable-stop-using-the-avx-avx2-instructions-in-drivers-/
Quote:
The 256-bit AVX registers are powered down. When a 256-bit AVX instruction is received the execution unit for 256-bit ops is powered on but there is some latency and this takes thousands of cycles to complete. During that time the 256-bit Ops are executed as a number of 128-bit Ops which can be 4 or 5 times slower. The AVX offset which technically is a bin limit, takes effect and stays in effect until 1 millisecond after the last 256-bit AVX instruction. Then the multiplier can return to it's normal bin.Originally Posted by chris719
Seems like it should switch to the AVX offset multiplier once AVX instructions are fetched from the instruction cache and bring itself out of the AVX offset multiplier as soon as the AVX instructions are done executing.
This means using just a handful of AVX 256-bit instructions can result in lower performance.
To check if the AVX offset is happening because of a bug or because there really are 256-bit AVX instructions then maybe one could try to see what happens if the Windows "BCDEdit /set xsavedisable 1" command is used but be aware if the software is not properly checking for OS AVX support there can be some crashes.
An example of a user wanting to try disabling AVX for other reasons, namely the bclk exploit in Skylake of which one of the effects is to never power on the 256-bit registers.
https://forums.geforce.com/default/topic/949301/how-to-disable-stop-using-the-avx-avx2-instructions-in-drivers-/
Quote:
Note also that an nVidia driver itself can be executing 256-bit AVX, it does not have to be the game itself.Originally Posted by AlphaC
"CPU runs AVX clocks if AVX offset is set, despite no AVX instructions being used to our knowledge. Seems to happen happen a lot in games. "
https://forums.overclockers.co.uk/threads/8packs-no-bs-testing-intel-8700k-and-z370-with-delid-5ghz-and-above-available-from-ocuk.18795256/page-29
Based on this report, it looks like this is because of actual AVX instructions being executed, not a bug.
Based on this report, it looks like this is because of actual AVX instructions being executed, not a bug.
Quote:
.
To individually adjust cores would require per core p-state control to be enabled and perhaps the 8700k does not have that. If so that could mean the CPU would be doing the 3 Musketeers and impacting all cores. Someone needs to test.
I would suspect your 7820X does have it though, is it enabled? With per core p-state enabled it certainly could do it on a per core basis but wether Intel set it up to do it that way I don't know. I'm still in the dark ages with older generations
![frown.gif]( "frown.gif")
Perhaps you could test it yourself.
As well as your AVX2 offset what is your AVX512 offset?
Would depend on the chip and setup. For instance the older HSW-E has an offset that is not settable as far as I know but seems to operate with the voltage table and number of active cores and dependent on the microcode patch used (avx offset was an after thought maybe).Originally Posted by TahoeDust
So, when an avx instruction is encountered does it just offset the clocks of the core the instruction is being run on?
.
To individually adjust cores would require per core p-state control to be enabled and perhaps the 8700k does not have that. If so that could mean the CPU would be doing the 3 Musketeers and impacting all cores. Someone needs to test.
I would suspect your 7820X does have it though, is it enabled? With per core p-state enabled it certainly could do it on a per core basis but wether Intel set it up to do it that way I don't know. I'm still in the dark ages with older generations
As well as your AVX2 offset what is your AVX512 offset?
After hearing 8700k owners complain about PUBG and other games triggering their AVX offset last night, I did this little test and that is not what I am seeing with my 7820x x299 setup. Sorry for the crudeness of the video, but you can see the AVX offset only looks like it is triggered once, very briefly at ~0:13. In this video my overclock is set to 4.8GHz with a AVX offset of -6. That is a little bigger offset than I normally run, but I wanted it to be easily observed.
So, what do we think is causing the bug in Coffee Lake? Bios? Microcode?
So, what do we think is causing the bug in Coffee Lake? Bios? Microcode?
Did you disable per core p-state to simulate an 8700k? It's easier to see what's happening if one checks the ratio and doesn't round it off to the nearest integer. For instance an average ratio of 47.6 over one or half a second might be rounded up to 48 with a lot of software. Also might consider disabling C1E, CC3, CC6, CC7 and speedstep so ratio is kept to max. If you feel like digging a little deeper then firestarter can apply AVX2 / AVX512 loads that are fairly constant but might want to use a lower max ratio for safety due to the high power draw. The loads can be applied as a duty cycle too. Also seems a little like it's still in beta stage, at least the w64 executable and of course use at your own risk.
Without a CFL chip I cannot make any judgment but it appears it's AVX activated and likely affecting all cores even if it only happens on one core. Perhaps the time it's activated has been extended, again needs someone to test. Is AVX2 considered a problem on CFL and designed to work this way, is it just some poorly selected setting or is it just a way of having some segmentation from other chips. I don't know, might be something entirely different.
One thing though, seems AVX is becoming a little more used, don't know when nVidia started using it in it's software but perhaps all those people who thought AVX wasn't a real world test and OC'd without taking it into consideration might get a rude awakening if newer drivers start crashing their system.
Without a CFL chip I cannot make any judgment but it appears it's AVX activated and likely affecting all cores even if it only happens on one core. Perhaps the time it's activated has been extended, again needs someone to test. Is AVX2 considered a problem on CFL and designed to work this way, is it just some poorly selected setting or is it just a way of having some segmentation from other chips. I don't know, might be something entirely different.
One thing though, seems AVX is becoming a little more used, don't know when nVidia started using it in it's software but perhaps all those people who thought AVX wasn't a real world test and OC'd without taking it into consideration might get a rude awakening if newer drivers start crashing their system.
Something else to also consider maybe.
Possibly the likes of CPU-Z and HWiNFO are not taking the average ratio but a short near instantaneous one which might happen during a brief AVX ratio window. HWiNFO settings can be changed regarding ratio calculation which in this case gets it closer but still appears off by a bit.
This is a bug. I tested different thread counts for Prime95 and anything below 5-6 threads will trigger the AVX offset even with a NON AVX prime95 load. Most games & apps are not using more than 4 threads, so the AVX offset is triggered all the time.
Just did a lot of testing and this bug is triggered when C-states are enabled. EIST does not affect it.
Enabling C1E will also trigger the bug but only when less than 6 threads are loading the CPU.
The CPU will still idle @ max multiplier when in high performance power plan setting, but the multiplier DROPS when a load is applied. This is clearly buggy behavior, the multiplier should never drop when a non AVX load is applied.
Just did a lot of testing and this bug is triggered when C-states are enabled. EIST does not affect it.
Enabling C1E will also trigger the bug but only when less than 6 threads are loading the CPU.
The CPU will still idle @ max multiplier when in high performance power plan setting, but the multiplier DROPS when a load is applied. This is clearly buggy behavior, the multiplier should never drop when a non AVX load is applied.
@HKPolice did you try "BCDEdit /set xsavedisable 1" ? Does the same thing happen?
Are you using W10? Does it still happen on W7 SP1 with Microsoft drivers?
What software are you using to measure clock ratio?, would you mind trying RealTemp GT or the i7 Turbo program.
Are you using W10? Does it still happen on W7 SP1 with Microsoft drivers?
What software are you using to measure clock ratio?, would you mind trying RealTemp GT or the i7 Turbo program.
Joined
·
1,986 Posts
Quote:
What are you using to monitor everything like that through your tablet?Originally Posted by TahoeDust
After hearing 8700k owners complain about PUBG and other games triggering their AVX offset last night, I did this little test and that is not what I am seeing with my 7820x x299 setup. Sorry for the crudeness of the video, but you can see the AVX offset only looks like it is triggered once, very briefly at ~0:13. In this video my overclock is set to 4.8GHz with a AVX offset of -6. That is a little bigger offset than I normally run, but I wanted it to be easily observed.
So, what do we think is causing the bug in Coffee Lake? Bios? Microcode?
Joined
·
55 Posts
Quote:
Please let us know !
Quote:
Please let us know !
Yeah, I'd really like to know how the computer was being monitored by a tablet too !!Originally Posted by SOCOM_HERO
What are you using to monitor everything like that through your tablet?
Please let us know !
Quote:
Yeah, I'd really like to know how the computer was being monitored by a tablet too !!Originally Posted by TahoeDust
After hearing 8700k owners complain about PUBG and other games triggering their AVX offset last night, I did this little test and that is not what I am seeing with my 7820x x299 setup. Sorry for the crudeness of the video, but you can see the AVX offset only looks like it is triggered once, very briefly at ~0:13. In this video my overclock is set to 4.8GHz with a AVX offset of -6. That is a little bigger offset than I normally run, but I wanted it to be easily observed.
So, what do we think is causing the bug in Coffee Lake? Bios? Microcode?
Please let us know !
Joined
·
3,757 Posts
Quote:
I see AVX affecting individual cores on through your hole Video. It's not a bug some games have AVX. From my testing AVX offset is activated when there is AXV data from the OS and games also Chrome.Originally Posted by TahoeDust
After hearing 8700k owners complain about PUBG and other games triggering their AVX offset last night, I did this little test and that is not what I am seeing with my 7820x x299 setup. Sorry for the crudeness of the video, but you can see the AVX offset only looks like it is triggered once, very briefly at ~0:13. In this video my overclock is set to 4.8GHz with a AVX offset of -6. That is a little bigger offset than I normally run, but I wanted it to be easily observed.
So, what do we think is causing the bug in Coffee Lake? Bios? Microcode?
any fix in the pipeline for this yet? Ive been digging as deep as i can and the most recent info is roughly a month old.
Just discovered this one on my new build myself. Im able to run 5gz w/ a 2 offset at llc5 and 1.32 (1.28 @ load) or (due to bug) im stuck running an offset of 0 which to get prime stable leaves me stuck using 1.29 w/ LLC6 which pumps me up to 1.344 in during the worst of it.
I can probably do the same on LLC5 by pumping my volts closer to 1.4 than id like - this pc has to last longer than most after we just bought a new roof ><
anyways - whats the word on the avx bug?
Just discovered this one on my new build myself. Im able to run 5gz w/ a 2 offset at llc5 and 1.32 (1.28 @ load) or (due to bug) im stuck running an offset of 0 which to get prime stable leaves me stuck using 1.29 w/ LLC6 which pumps me up to 1.344 in during the worst of it.
I can probably do the same on LLC5 by pumping my volts closer to 1.4 than id like - this pc has to last longer than most after we just bought a new roof ><
anyways - whats the word on the avx bug?
Joined
·
34 Posts
Having the same problem
I'm having the same problem on my i7 6950x overclocked to 4.2/4.1 with avx offset of 2.
When I'm playing PUBG or just using Google Chrome I noticed that some cores clock down like it would with AVX. I tried all that I could to fix this but no sucess... I guess that chrome and some games like PUBG are using AVX now.
I tried to use the command line: "bcdedit /set xsavedisable 1" to disable AVX but this crashed my Corsair Headset software and driver, so I had to enable it again.
Anyone have any idea if chrome and games are really using AVX or if this is a bug?
Thank you all!
I'm having the same problem on my i7 6950x overclocked to 4.2/4.1 with avx offset of 2.
When I'm playing PUBG or just using Google Chrome I noticed that some cores clock down like it would with AVX. I tried all that I could to fix this but no sucess... I guess that chrome and some games like PUBG are using AVX now.
I tried to use the command line: "bcdedit /set xsavedisable 1" to disable AVX but this crashed my Corsair Headset software and driver, so I had to enable it again.
Anyone have any idea if chrome and games are really using AVX or if this is a bug?
Thank you all!
Joined
·
3,757 Posts
Chrome and some games use AVX also some background processes in the OS.
1 - 20 of 21 Posts
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
-
?
