Low memory clock on Maxwell2 cards (960/970/980, probably Titan X)

archae86
archae86
Joined: 6 Dec 05
Posts: 2,673
Credit: 2,353,645,943
RAC: 3,015,929

I stumbled on Nvidia

I stumbled on Nvidia documentation for the GPU Performance State Interface

While one comment expressly says "The GPU performance state APIs are used to get and set various performance levels on a per-GPU basis" my initial scan saw lets of get and no set.

Still, possibly in here is a hint at how applications may influence the performance state.

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 758
Credit: 153,316,509
RAC: 137,456

@Harri: wow, 80 - 85% MCU

@Harri: wow, 80 - 85% MCU load! It looks like SETI would benefit even more from higher memory clocks.

Regarding the driver profiles: the only option which I identify as possibly relevant is the power management mode. However, trying all setting I could not provoke any changes running GPU-Grid.

@archae: Maybe nVidia deprecated the ability to force the P-state manually? This would explain why there's nos et function and why the things I found on the net (mostly written for older drivers & cards) don't have any effect.

MrS

Scanning for our furry friends since Jan 2002

DanNeely
DanNeely
Joined: 4 Sep 05
Posts: 1,281
Credit: 1,414,128,522
RAC: 1,189,908

My GeForce 770, currently

My GeForce 770, currently running 3x BRP4 WUs, is at P0, 89-96% GPU load, 52-59% MCU load.

When I opened the overclocking section, the only P options I had on the dropdown were P0/1/5/8; no P2.

I'm using driver v344.11.

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 758
Credit: 153,316,509
RAC: 137,456

Thanks, but your GTX770 is a

Thanks, but your GTX770 is a Kepler, not a Maxwell. The problem doesn't apply to them. Maybe I should have been more clear about this.

MrS

Scanning for our furry friends since Jan 2002

disturber
disturber
Joined: 26 Oct 14
Posts: 30
Credit: 57,155,818
RAC: 0

I noticed that my memory

I noticed that my memory clock was not running at the speced rate of 3505 MHz a while back. Using the Gigabyte OC guru would not let me change the frequency at all. As soon as i pressed apply, it would jump back to the lower clock rate. It appears that you cannot enable the P0 mode when running compute tasks.

My gpu load using a i7-3770k running at 4.4GHz is at 96% running at two BRP4G tasks. The memory controller load is around 75% and each WU fishes in 54 minutes on the average.

Searching for the p2 state brought me to the bitcointalk.org mining forum. The solution they suggested also worked for me.

Download NVIDIA Inspector 1.9.7.3., Run it and show the overclocking tab. Select the P2 performance level, P0 is greyed out. Select unlock max and then chose a Memory clock frequency you are comfortable with. I kept adding 20 MHz at a time until I reached 3504, the clock rate the card is rated for without any issues. I could not increase any voltages because they were greyed out. Unless this produces work unit errors, I will keep it at that and report back the improvements.

I did no see any temperature or power increase. Both are 55C and 49% respectively.

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 758
Credit: 153,316,509
RAC: 137,456

Disturber, thanks for your

Disturber, thanks for your report. I thought my section "Increasing memory clock speed:" included the same info, but may have lacked detail. Let us know if this setup is stable for you. As I stated it's not stable for me, although it works just fine in P0.

MrS

Scanning for our furry friends since Jan 2002

archae86
archae86
Joined: 6 Dec 05
Posts: 2,673
Credit: 2,353,645,943
RAC: 3,015,929

RE: I did no see any

Quote:
I did no see any temperature or power increase. Both are 55C and 49% respectively.


It would be surprising if the actual useful work output increased with neither an increase in temperature or power.
Unless, of course, the fan rate went up, or the power reported by the software link did not correspond to the real change in power consumption.

Let us know what you see in terms of actual unit completion elapsed time.

disturber
disturber
Joined: 26 Oct 14
Posts: 30
Credit: 57,155,818
RAC: 0

I have the results so

I have the results so far.

All the work units run slightly less than 50 min apiece and validate without errors. This works out to roughly 10% gain. I am not willing to try it any higher right now, the gpu core clock runs at 1392 MHz just fine.

http://einsteinathome.org/host/11685173/tasks&offset=40&show_names=1&state=0&appid=25

The card I have is the Gigabyte GTX 970 G1 Gaming card, so maybe the memory chips are binned to a higher spec? They seem to work fine at 3506 MHz (7012 MHz data rate).

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 758
Credit: 153,316,509
RAC: 137,456

A 10% gain is very nice for a

A 10% gain is very nice for a 16.7% memory frequency increase, especially if it didn't increase the power draw! You previously reported a memory controller load of 75% - did this decrease?

Quote:
The card I have is the Gigabyte GTX 970 G1 Gaming card, so maybe the memory chips are binned to a higher spec? They seem to work fine at 3506 MHz (7012 MHz data rate).


Well, the memory has to work at 7.0 GHz data rate according to the cards specification ;)

MrS

Scanning for our furry friends since Jan 2002

disturber
disturber
Joined: 26 Oct 14
Posts: 30
Credit: 57,155,818
RAC: 0

memory controller load jumps

memory controller load jumps a lot

at 3005 MHz:
MCU load: 72-83%
power: 45-48% TDP

at 3506 MHz:
MCU load: 70-78%
power: 48-50% TDP

So power went up slightly and MCU went down a bit.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.