Overclock.net banner
3061 - 3080 of 3091 Posts

·
Registered
Joined
·
136 Posts
For those looking to use adaptive to undervolt their 10980xe, the best frequency to do that is at 2.1 GHz using AVX-512.

That is really one setting to rule them all because the VID for AVX-512 will not start scaling under 2.2 so we have to make sure 2.1 is stable.

The reason to use AVX-512 even if you don’t care about AVX-512 is because SSE started to scale at 1.3 so if you are using SSE then AVX-512 will give you instability at stock or even below stock.

Do not afraid to use p95 with AVX-512 since at 2.1 it will draw very low power.

Shaved 0.075v, worth 1 or 2 multiplier depends on how high is the OC.
 

·
OC...D
Joined
·
28,239 Posts
Hey gang -

I'm a year or two late to this thread it looks. I just picked up a 10980XE from Microcenter last week and have been working on OC'ing it. I have a great deal of ignorance about all of the various bits that can be tweaked. Primarily I'm just used to pushing the core ratio, any AVX offsets needed, and Vcore values. I'm replacing a Silicon Lottery binned and delidded 7900X which could comfortably hit 4.7Ghz with a 3/5 offset and 1.237 Vcore. The unofficial goal of the 10980XE was to see if I could hit 5GHz all-core and keep it cool enough.

Uh. Sorta. Initial findings see a 1.367(!) Vcore will keep the chip from crashing CPU intensive applications at 5GHz. Anything lower, and it's crash city. After 10-20 minutes of gaming, the temps hit in the low-mid 80s*C, with a dedicated water cooling loop. Now I know the EK R6E monoblock I have isn't perfectly ideal, and I'm looking to replace it. But I'll bet that's only good for a few degrees C.

Dropping the speed to 4.8GHz sees a required Vcore of 1.250. That's pretty substantial, and easily drops the gaming temps by 10 or so degrees. Ultimately I'm not sure that 200MHz is going to make a huge difference when fragging, but after all of that, here's my question:

Is a 1.367 Vcore actually bad for the CPU? Is it just a case of struggling with cooling? Or are there other things I'd have to keep aware of?

Specifics:
  • Cooling path: Pump --> EK R6E monoblock --> AlphaCool 480x60mm radiator in a push/pull setup --> reservoir/pump.
  • Idle temps drop right down to 40*C and stay there.
  • Gaming temps hit low 80s*C at 5Ghz. 70-75*C at 4.8GHz.
Any thoughts? Thanks!
A noble objective, but a 24/7 5.0 10980XE is going to require direct-to-die cooling. A lower core-count CPU would have been a better choice for 5.0 with most games.

You will find that once you tune the ram to the 4000c16 range, and cache to 30X + 4.7 to 4.8GHz on these chips they work fine in CPU-intensive gaming. On the other hand, doing a per-core with 2 cores at 5.0 and the rest at 4.8 may be your best target OC. Adaptive vs Static voltage really affects littlie in terms of life-span and cooling ability, since the power draw at idle is nominal for both (tho you will note that when using adaptive, the idle amperage is much higher than when using static voltage... as it should be :) )
Start by getting a rock solid 4.7 OC, tune the ram and cache, then see if your cooling kit can manage to keep the temperatures under 80C during gaming (not peak temps, but the median or average). Dial the core OC up until the at load temperatures are not much more than 80C. Note Elmor's advice in the OP... VCCIN is key to controlling the temperatures. And always allow for some healthy droop in VCCIN at load. Disable VCCIN tracking, disable speedstep and ENABLE speedshift for the best clock response times. 🤙
Give this article a read. It's not simply the voltage at load that affects these HCC chips, it is the power draw and temperature at voltage that burns the fuse...
 

·
Registered
Joined
·
46 Posts
A noble objective, but a 24/7 5.0 10980XE is going to require direct-to-die cooling. A lower core-count CPU would have been a better choice for 5.0 with most games.
Perhaps. Too late now. :) Time to do the best I can with what I have here. I've backed it off from the 5.0 to 4.8, with a 1.250 Vcore. Seems to keep the gaming temps in the 60s, sometimes kissing 70*.

You will find that once you tune the ram to the 4000c16 range, and cache to 30X + 4.7 to 4.8GHz on these chips they work fine in CPU-intensive gaming.
RAM is and has been running at 4GHz since I bought it 4 years ago. So I'm good there. If I try to push it to its max of 4.2GHz, I never even see a BIOS screen. Good times.

The cache tuning is something I'm 100% ignorant of. Could use some help understanding that.

On the other hand, doing a per-core with 2 cores at 5.0 and the rest at 4.8 may be your best target OC.
OK, with this one, would I set the voltage of each core to "Auto" and let the BIOS try and figure it out? Or use my two known, working Vcore numbers of 1.367 and 1.250 for the 5.0 vs 4.8GHz cores respectively? From what I've seen, the BIOS autos way too high. Like when I, for S'n'Gs ran the 5GHz all-core on Auto, it set the Vcore to 1.4052.

Note Elmor's advice in the OP... VCCIN is key to controlling the temperatures. And always allow for some healthy droop in VCCIN at load. Disable VCCIN tracking, disable speedstep and ENABLE speedshift for the best clock response times.
I have to admit that all went way over my head. Again: all I'm used to playing around with are the ratios and Vcore. Any suggestions on this?

Appreciate it.
 

·
OC...D
Joined
·
28,239 Posts
Perhaps. Too late now. :) Time to do the best I can with what I have here. I've backed it off from the 5.0 to 4.8, with a 1.250 Vcore. Seems to keep the gaming temps in the 60s, sometimes kissing 70*.

RAM is and has been running at 4GHz since I bought it 4 years ago. So I'm good there. If I try to push it to its max of 4.2GHz, I never even see a BIOS screen. Good times.

The cache tuning is something I'm 100% ignorant of. Could use some help understanding that.



OK, with this one, would I set the voltage of each core to "Auto" and let the BIOS try and figure it out? Or use my two known, working Vcore numbers of 1.367 and 1.250 for the 5.0 vs 4.8GHz cores respectively? From what I've seen, the BIOS autos way too high. Like when I, for S'n'Gs ran the 5GHz all-core on Auto, it set the Vcore to 1.4052.



I have to admit that all went way over my head. Again: all I'm used to playing around with are the ratios and Vcore. Any suggestions on this?


Appreciate it.
Yep, okay - for the per-core. On the main bios voltage page you set this to Auto Vcore, then scroll up and where you have "All Core" selected, select "per specific core" or per individual core (depending on which bios is on that R6E). Then enter the menu bringing you to the voltage and multiplier setting for each of the 18 cores. Two cores will have an * ... these are the cores with the "most favorable" Hz/mV curve from during VID programming (the least leaky?). Use manual override for each of the 16 other cores @ 1.25V for 4.8. Then for the * cores, set these to 49 or 50 and the voltage needed for that frequency 1.37V ?). I have a bios dump of a per core somewhere (in this or the Asus x299 thread...)

Ram - have you stress tested the ram at 4000? (GSAT or TM5?)

Cache - there's a point of diminishing returns 30-32 with the Vcache below 1.25V is the sweetspot on this platform. Keep an eye out for correctable WHEA in the system logger - they kill efficiency.

VCCIN - one way to get at the min vccin needed for any given set of frequencies (or a single all core frequency) is to use something like R20 or Blender and start lowering the VCCIN (but keep some vdroop active) until you see a decline in performance (score) or it hangs/crashes. Then up vccin to the last known glitch-free value. Then press the "easy" button! 😉
 

·
Registered
Joined
·
60 Posts
A noble objective, but a 24/7 5.0 10980XE is going to require direct-to-die cooling. A lower core-count CPU would have been a better choice for 5.0 with most games.

You will find that once you tune the ram to the 4000c16 range, and cache to 30X + 4.7 to 4.8GHz on these chips they work fine in CPU-intensive gaming. On the other hand, doing a per-core with 2 cores at 5.0 and the rest at 4.8 may be your best target OC. Adaptive vs Static voltage really affects littlie in terms of life-span and cooling ability, since the power draw at idle is nominal for both (tho you will note that when using adaptive, the idle amperage is much higher than when using static voltage... as it should be :) )
Start by getting a rock solid 4.7 OC, tune the ram and cache, then see if your cooling kit can manage to keep the temperatures under 80C during gaming (not peak temps, but the median or average). Dial the core OC up until the at load temperatures are not much more than 80C. Note Elmor's advice in the OP... VCCIN is key to controlling the temperatures. And always allow for some healthy droop in VCCIN at load. Disable VCCIN tracking, disable speedstep and ENABLE speedshift for the best clock response times. 🤙
Give this article a read. It's not simply the voltage at load that affects these HCC chips, it is the power draw and temperature at voltage that burns the fuse...
Hi, I really thought the current difference between adaptive and static is negligible at idle. Not enough to start worrying about it.

When the CPU idle, running static still result at higher power draw which means the current draw difference is not linear.
Now assuming current flow at static is 50W / 1. 2v = 42A, with adaptive say it idle at 35W / 0.7v = 50A so it resulting in 8A difference which I believe is negligible compared to +200A under load.

I have different opinion about VCCIN too.
At high power draw, the power plane and traces between the VRM and the CPU FIVR (this includes board traces and CPU traces) will start to heats up due to high current flowing and it will heat up the CPU more than the heat from higher VCCIN.

Only way to reduce the heat from the power plane so that it won’t affects your CPU temp without lowering the OC is to supply more VCCIN. I don’t say that we should shove dangerous amount of VCCIN but what I’m trying to say is lowering VCCIN as much as possible is counterproductive in my opinion.

It sounds bizzare (it does) but I can prove it and actually anyone also can easily compare their VRM temp between lower and higher VCCIN themselves. Note that the VRM temp doesn’t affects the CPU temp directly but the heat from the power plane will especially the power plane inside the CPU between the pads and the FIVR.
 

·
Registered
Joined
·
60 Posts
adaptive is basically unusable on the 10 series cpus so not really even worrying about it lol.
Lol yeah, but it could be useful depends on how high your OC is and your VID. Plus with some degree of negative offset you can push it a bit further.
 

·
Registered
Joined
·
46 Posts
Yep, okay - for the per-core. On the main bios voltage page you set this to Auto Vcore, then scroll up and where you have "All Core" selected, select "per specific core" or per individual core (depending on which bios is on that R6E).
OK, that makes sense and is what I tried to do. Crash city. I backed off to my formerly stable 4.8GHz / 1.250 on all cores, and... crash city. :-( I'm now up to 1.259 Vcore for a stable 4.8. Looks like I'm suffering the voltage degradation others have mentioned. Hopefully haven't borked the CPU.

Ram - have you stress tested the ram at 4000? (GSAT or TM5?)
No need. It's a gaming rig after all.

Cache - there's a point of diminishing returns 30-32 with the Vcache below 1.25V is the sweetspot on this platform. Keep an eye out for correctable WHEA in the system logger - they kill efficiency.
I was trying to play around with this and got nothing but crash city. So reset it all to "Auto" and left it alone.

VCCIN - one way to get at the min vccin needed for any given set of frequencies (or a single all core frequency) is to use something like R20 or Blender and start lowering the VCCIN (but keep some vdroop active)
Vdroop makes 0 sense to me. No idea what you're referring to there.
 

·
OC...D
Joined
·
28,239 Posts
OK, that makes sense and is what I tried to do. Crash city. I backed off to my formerly stable 4.8GHz / 1.250 on all cores, and... crash city. :-( I'm now up to 1.259 Vcore for a stable 4.8. Looks like I'm suffering the voltage degradation others have mentioned. Hopefully haven't borked the CPU.
No need. It's a gaming rig after all.
I was trying to play around with this and got nothing but crash city. So reset it all to "Auto" and left it alone.
Vdroop makes 0 sense to me. No idea what you're referring to there.
You are most likely seeing the CPU "round out" (etchings) causing the very slightly higher vcore for 4.8. It's common and to be expected. I've attached a bios text file with a simple [email protected] [email protected] per core. It may help you.
Vdroop is controlled by the LLC (load line compensation). In any DC circuit at "constant voltage" (adaptive or manual override in our case) when the current/load changes, the voltage must undergo an oscillation around the set value. This happens on the uSec scale and is not anything you can see without special tools (reading from the socket). MOdern VRMs are quite good at controlling for this, but frankly the overshoot in Voss is the most common cause of degradation and Intel specifies limits for this transient Load Line Overshoot in their spec docs to board engineers. So... the BEs introduced Load Line Compensation (LLC) back in the socket 775 days to address this failure source. (extreme benchers were seeing a bench fail when the the benchmark was ending - load line undershoot from that oscillation around the set voltage).
On your board, you should set an LLC value of 5 (mid range) and adjust VCCIN if you see an undervolt condition during heavy loads. In the pics below, "offset" = load line voltage compensation. Lol, be advised... the last thing an OCer should do is attempt to meet the voltage requirement of the chip at load by defeating the board's LLC fun=ction (eg, on Asus, a value of 8 or higher.
At load you should see the VCCIN "droop" = drop to a lower value. If you see it inccrease under load: Danger Will Robinson. :)

As far as checking the ram... it can be a source of correctable errors = lower performance as the proc_call will repeat in an attempt to match checksums, and if it can't after a few repeats (loss of efficiency) then you have an uncorrectable WHEA and a "Halt" command will be issued by the OS. And in my experience, bad ram settings is the only way you can actually get windows to forget it's name! :)

But, as I always say... Smoke 'em if you got 'em. 🖖

transient_no_vdroop.jpg transient_no_vdroop_no_offset.jpg
 

Attachments

·
OC...D
Joined
·
28,239 Posts
Hi, I really thought the current difference between adaptive and static is negligible at idle. Not enough to start worrying about it.

When the CPU idle, running static still result at higher power draw which means the current draw difference is not linear.
Now assuming current flow at static is 50W / 1. 2v = 42A, with adaptive say it idle at 35W / 0.7v = 50A so it resulting in 8A difference which I believe is negligible compared to +200A under load.

I have different opinion about VCCIN too.
At high power draw, the power plane and traces between the VRM and the CPU FIVR (this includes board traces and CPU traces) will start to heats up due to high current flowing and it will heat up the CPU more than the heat from higher VCCIN.

Only way to reduce the heat from the power plane so that it won’t affects your CPU temp without lowering the OC is to supply more VCCIN. I don’t say that we should shove dangerous amount of VCCIN but what I’m trying to say is lowering VCCIN as much as possible is counterproductive in my opinion.

It sounds bizzare (it does) but I can prove it and actually anyone also can easily compare their VRM temp between lower and higher VCCIN themselves. Note that the VRM temp doesn’t affects the CPU temp directly but the heat from the power plane will especially the power plane inside the CPU between the pads and the FIVR.
So if I understand what you are positing... you have a scenario where increasing voltage on the VCCIN rail lowered CPU temperature under load? 🤨
 

·
Registered
Joined
·
60 Posts
So if I understand what you are positing... you have a scenario where increasing voltage on the VCCIN rail lowered CPU temperature under load? 🤨
Hi there, yes. I know it sounds weird but it is true and it is actually make sense following the V = I x R equation.

I was having issue with hot board traces between vrm and the cpu socket and I thought if the traces is not enough then I probably can increase the voltage so it doesn’t need that much current flow under same power draw. And it actually worked (still hot as the power draw is high with high oc of 4.7 but not as hot as before).

Then I start to realize that my CPU temp is much lower than before and true enough lowering VCCIN increased it’s temp back to previous. It won’t appear under light load where the current draw is not enough to heat up the power plane between the CPU pads and the FIVR. But once the current draw is high (stress test such as occt or high oc + cb23) then it start to happen and make difference. This is with VIN Tracker disabled, with it enabled the cpu simply throttle so can’t see the effects.
 

·
OC...D
Joined
·
28,239 Posts
Yeah, the issue is that the CPU uses a (tracking) down converter to to adjust the VCCIN to the necessary on-die voltages - this could be the heat source. But the effect you observed (heat sinking to the socket) would, in the voltage scenario you describe show more as heat soaking on the VRMs from the power plane. Either way, Ohm's law manifesting as heat (from power demands for the work requested) could be due to a single flaw/imperfection (choke point) in the tracings in that specific board. On 4 x299 boards here, I can't reproduce the effect since you posted it some time ago.

I've had this 10980xe at 4.8 for most times and 4.7 for CPU intensive work since the SKU launched (and a few excursions in the 5.0-5.2 range for benching) and using an IR thermometer, I do not see the PCB heating on this R6EO or the R6Apex. Haven't seen it with the delided 7980XE on the Apex. Not tried it on the Giga G9, or Designare.
 

·
Registered
Joined
·
60 Posts
Yeah, the issue is that the CPU uses a (tracking) down converter to to adjust the VCCIN to the necessary on-die voltages - this could be the heat source. But the effect you observed (heat sinking to the socket) would, in the voltage scenario you describe show more as heat soaking on the VRMs from the power plane. Either way, Ohm's law manifesting as heat (from power demands for the work requested) could be due to a single flaw/imperfection (choke point) in the tracings in that specific board. On 4 x299 boards here, I can't reproduce the effect since you posted it some time ago.

I've had this 10980xe at 4.8 for most times and 4.7 for CPU intensive work since the SKU launched (and a few excursions in the 5.0-5.2 range for benching) and using an IR thermometer, I do not see the PCB heating on this R6EO or the R6Apex. Haven't seen it with the delided 7980XE on the Apex. Not tried it on the Giga G9, or Designare.
Actually if we are looking at the x299 VRM as power supply to the FIVR in the same way we are looking at the PSU to the VRM on mainstream platform then it make sense not to starve the VCCIN in the same manner that we will not lower the 12 v rail on mainstream platform.

I even believe if we lower the 12 v rail on mainstream platform up to certain threshold it will still work without issue since it is down converted to much lower voltages. While for x299 VCCIN it is only few hundreds milivolts difference with the actual usage (vcores etc depends on your OC) which can easily swing the result.
 

·
OC...D
Joined
·
28,239 Posts
No problems here, this 10980XE/R6EO has been fine at [email protected] 1.185V with 1.85V VCCIN LLC5 for a very long time. :)
 

·
Old and Decrepit
Joined
·
2,314 Posts
adaptive is basically unusable on the 10 series cpus so not really even worrying about it lol.
Thats a fact jack! Adaptive offset cannot be lower than the VID which is outragous. I about crapped myself on my first 4.8 run right after installing the chip. just set it to 48 all cores and the Vcore shot up ti 1.414! Ive tried multiple methods and what works best for me is a static Vcore. With Intel speed shift and turbo enabled. If I disable the turbo I get 3.0 all cores no matter what I set the multiplier to. You dont need turbo installed, just the driver.
 

·
Registered
Joined
·
3 Posts
I've got to ask, how do you overclock a 10980XE if Adaptive is so broken, I'm trying to make a daily OC but can't really do it because of that so my temps end up far worse than everyone elses
 

·
OC...D
Joined
·
28,239 Posts
I've got to ask, how do you overclock a 10980XE if Adaptive is so broken, I'm trying to make a daily OC but can't really do it because of that so my temps end up far worse than everyone elses
Use Manual override and disable CPU SVID. You'll find that the voltage necessary for 4.7 or 4,8 (reasonable overclocks) is many times a lot lower than the VID. As an example, and there are many in this thread. My 10980XE is stable and easily temp controlled at 4.7 with 1.185V and 1.85VCCIN LLC5. Since Adaptive can only add voltage to the VID the min requested vcore (=VID) for this chip is over 1.4V for 4.7GHz. It's not the board(s)... it is the poorly programmed VID.
Static voltage at idle clocks and temperature is meaningless/harmless... within reason, of course.
 

·
Old and Decrepit
Joined
·
2,314 Posts
^^^^
This is exactly what I run 24x7. I can get higher but it requires more Vcore which results in more heat. The difference from 4.7 to 4.8 for me is 1.185Vcore @4.7 and 1.225 @ 4.8. Loaded temps go up but that extra 100Mhz doesn't change the performance by much. Id rather stay in the cool 70s than to gain 400 points in CB23. Blender renders AVX workload with no AVX offset are a few seconds faster but not enough to make a difference, only higher temps.
 

·
Registered
Joined
·
136 Posts
My DIY wooden desktop PC finally finished!

2517586


2517587


BTW regarding adaptive, if you are using AVX-512, you are lucky if you can even get -10mv.

But in the other hand adaptive without offset (running Intel default VID) is actually the only way if you are expecting high AVX-512 speed.

With manual voltage it will be too much for AVX-512 if your SSE is at 46 and above and your SSE will be too low if you aim manual voltage for AVX-512.

I myself running adapyive without offset for 14x4.6, 2x4.7 and 2x4.8. That way I can run my AVX-512 with -8 offset and stable on prime95 small FFTs and linpack. Actually linpack while drawing less wattage than p95, it is easier to detect vcore instability because it switched between SSE and AVX-512 constantly.
 

·
Registered
Joined
·
3 Posts
Cheers, works fine at 4.7GHz 1.180v static, really strange, why even include adaptive as a setting if it only works when overvolting...
my next question is, What are your cooling solutions like? as when i ran tests with said settings i was having strange temps like 65° to 95° core temps, now my setup shouldn't be having trouble as its got 2x 480mm XT45's and an EK motherboard block (R6EE) and I've repasted twice to see the same cores 30 degrees hotter than the coldest cores.
Safe to say i've no idea what i'm even doing anymore :)
 
3061 - 3080 of 3091 Posts
Top