CPU Researching 'performance-per-watt' and power efficiency? (Undervolting, etc.)

[Chrizz]

The RVZ01 has pretty good cooling for the GPU. Make sure you feed it with two 120mm fans and keep the case standing.

The case is standing, but I only have one case fan for the GPU. Now that I think of it, that fan is controlled by the CPU temperature, so when benchmarking with heaven or valley, this fan is not spinning and is more of an obstruction than a help.
The GPU I have is the Gigabyte GeForce GTX 970 4GB G1 Gaming, which should be one of the best cooling cards out there.

 

[Thehack]

The case is standing, but I only have one case fan for the GPU. Now that I think of it, that fan is controlled by the CPU temperature, so when benchmarking with heaven or valley, this fan is not spinning and is more of an obstruction than a help.
The GPU I have is the Gigabyte GeForce GTX 970 4GB G1 Gaming, which should be one of the best cooling cards out there.

You should get two case fans with decent static pressure and have them pump air into that chamber. If you have questions just send me a message. We don't want to muddy this thread too much.

 

[Raxe]

Sorry, I think I was a little unclear about this. You can click L to lock it to a specific frequency-voltage combination while you're testing what works well for your card. Then you can disable the lock when you're done. As long as your intended peak combination, in your case +300mhz at 0.85v, is the highest frequency over every subsequent voltage point, it will clock down and up along the curve from idle to that point. At least, that's what I found with my testing. I currently lock it because I'm always running primegrid.

Oh god, it's true, awesome. So it is like an mix of a "fixed and an offset" mode of overclocking the cpu. The best of two worlds. The only requirement to "destroy" the famous boost clock 3.0 in the pascal cards seems to be a linear curve when you want to "stop" the clocks and voltage go further.

When i return to home i need to do some testing about temperatures&consumption gains, if these specs on paper are true on reality this is going to be an awesome tool to fit high wattage cards with low power psus with afterburner.

 

[Biowarejak]

The case is standing, but I only have one case fan for the GPU. Now that I think of it, that fan is controlled by the CPU temperature, so when benchmarking with heaven or valley, this fan is not spinning and is more of an obstruction than a help.
The GPU I have is the Gigabyte GeForce GTX 970 4GB G1 Gaming, which should be one of the best cooling cards out there.

IIRC, EKWB has a cable to go from PWM graphics card fans to PWM case fans.

 

[Raxe]

Ok, so i´m going to post some of the "benchmarks" done with the new curve for voltage/frequency in Msi Afterburner. Before anything i don´t have and not pretend to buy a Kill a Watt, cause i´m spanish and in here Amazon list them for 153 and 174 euros respectively lol

So instead, i use hwinfo for power consumption.

Ah, and i´m using a custom w.i.p realtime character that i´m working on instead of a typical benchmark/ furmark extreme consumption, cause for me this is the real consumption for my workflow/pipeline archiving 100% usage on the gpu( and generating a lot less coil whine than using unigine valley).

My system:

-i7 3770k 3.521mhz ( Turbo disabled) w/ intel cooper base stock cooler
-Asrock h67m-itx
-16gb 1333mhz
-Palit dual gtx 1060( So toasty and lightweight)
-Crucial Mx300 275gb
-5 2.5hdd
-Custom Rvz02(accumulating old laptop hdd´s)
-Sfx-L silverstone 500w


The methodology consist of:

*Ambient Tº 11º C, luckily it´s a cold day.

*One scene, 4 assets, several textures from 1k to 8k, all the graphic options pushed to the max on purpose of using permanent 100% gpu.

*Few minutes of stress, cause i´ve said earlier, i want to have a real monitoring of my daily usage, not an 8 hours furmark melting my system without sense for me.

*Only Vivaldi and Franz opened as far as secondary programs active between the benchmarks.

*Capture of all the stats, power and voltages in Afterburner, Hwinfo, and the monitoring rivatuner, for avoiding "fakes".

*I´m not relaxing the gpu, so the iddle temperature so high in some captures( The first test mostly) is because it idles for 30 seconds or meanwhile i´m writting this post after the stress,so, not much time to cold down.

*5 Afterburner profiles, 1 stock. All the afterburner profiles share the same ventilation curve( Ok, it seems that test 1 was fuc..up with the fan curve, so expect 2-3 degrees higher).

For Average temperatures/watts/voltage see pictures:

TEST #1 STOCK . BOOST CLOCK 1.86Ghz + 0Mhz Mem, 1.05V, 70ºC Max [9.5 FPS] 106W MAX


TEST #2 VOLTAGE CURVE. BOOST DISABLED 1.708Ghz -200Mhz Mem, 0.8V, 55ºC [8.2 FPS] 59 W MAX


TEST #3 VOLTAGE CURVE. BOOST DISABLED. 1.911Ghz Fixed +350 Mhz Mem 0.9V, 65ºC Max [9.9 FPS] 84W MAX [POWER +16%]


TEST #4 VOLTAGE CURVE. BOOST DISABLED. 2Ghz Fixed +400 Mhz Mem 0.95V, 63ºC Max [9.7 FPS] 89W MAX


TEST #5 OC. BOOST CLOCK +210Mhz ( 2.075 Ghz) Core +500 Mhz Mem 1.063V, 75ºC Max [10.7 FPS] 125W MAX [POWER +16%]


As you can see, and i´m not finished optimizing and finding the best core/memory overclock for the voltage, you can shave a lot of power consumption/temperature/voltage killing the new Nvidia Pascal Boost 3.0,that is, locking the voltage at the freq you want with a linear constant curve.

And finally, see that ultralow % of fan curve allowed with this, i´m sure that is arround 1k rpm or less, so noise gratefulness.

PD: Personally i don´t need lower consumptions and i not have a super limited cooling for my gpu, but i hate coil whine( and with this i don´t ear soo much) and also i hate the noisyness/temperature of this palit 1060, in stock and with low fans(bellow 35% it´s easy to have 80-80´s mid in super stressful tasks in a hot day). And of course this is awesome for sff pc´s and brickless builds with the new HDPLEX, if it´s in reality the same as the values in paper.

If anyone have any suggestion or doubt about this, tell me.

 

[Raxe]

Investing some time on this, i think that i found my sweet spot in undervolting.

STOCK: 70ºC, 1.05V 105W

UNDERVOLT: 55ºC, 0.85V 70W

So, roughtly -35w and -15ºC for the same core mhz and some oc in the memory.

 

[alexep7]

It's just the specific "Kill-a-watt" brand that isn't for sale at decent prices there, I saw a regular watt meter on spanish amazon for 13€ just the other day

 

[zovc]

Thanks @Raxe! The difference between Stock (Test #1) and your first adjustment (Test #2) is tremendous! Test #3 and Test #4 are really interesting results, too.

I honestly like using HWiNFO64 or other software for monitoring wattage just because I feel like--even if it's inconsistent from card-to-card or board-to-board--I think it should be consistent within a given system. I would like for more people to mention (or at least include) software wattage monitoring if only to see what the computer thinks the part being tested is drawing. An outlet monitor like the popular Kill-a-Watt (which is only popular here because it's cheap, not because it's especially great or accurate) is always going to report your entire system usage and, while you can do your best to minimize other parts' fluctuations, they'll always be there to some degree.

 

[Raxe]

Thanks @Raxe! The difference between Stock (Test #1) and your first adjustment (Test #2) is tremendous! Test #3 and Test #4 are really interesting results, too.

I honestly like using HWiNFO64 or other software for monitoring wattage just because I feel like--even if it's inconsistent from card-to-card or board-to-board--I think it should be consistent within a given system. I would like for more people to mention (or at least include) software wattage monitoring if only to see what the computer thinks the part being tested is drawing. An outlet monitor like the popular Kill-a-Watt (which is only popular here because it's cheap, not because it's especially great or accurate) is always going to report your entire system usage and, while you can do your best to minimize other parts' fluctuations, they'll always be there to some degree.

Thanks @zovc
It´s a shame that this level of customization on voltages seems to be only for the new Pascals, but if i´m not wrong in memory, i remember than in the past years, nvidia inspector can adjust and lock voltages in a similar way, but more complex and tedious i suppose.

I was amazed when @Zeroth Alpha shows this working well and well, the results speaks for itself.

And about that massive difference of wattage between Test 1 & 2 it´s mostly the reduction of a whooping -0.25V, a lot of core clock and even -200Mhz to the memory. In fps it´s about a -20% performance, a big hit.

The last test, the "Sweet Spot" have a much better performance/watt for me, same performance as stock at a 2/3 the power.

In my case, anything bellow 0.95-0.9V increases power consumption in tiny bits, included memory overclocking, that adds arround 6-7W when overclocked +350mhz at 0.9V. If i put the 0.8V profile, memory undervolting to -200Mhz only shaves roughtly 1 W, it´s not worth it downclocking the memory at that low voltages

 

[Zeroth Alpha]

@Raxe Your testing is very comprehensive. I'm glad that you find it helpful. This is something I would definitely recommend doing for anyone with a new pascal card. The lower temperatures and quieter card is definitely worth it. For me, I personally monitor change in power consumption via the power limit graph in afterburner. I see a drop from around 90% to 65% at equivalent performance. Also a GTX1060.

 

[Raxe]

@Raxe Your testing is very comprehensive. I'm glad that you find it helpful. This is something I would definitely recommend doing for anyone with a new pascal card. The lower temperatures and quieter card is definitely worth it. For me, I personally monitor change in power consumption via the power limit graph in afterburner. I see a drop from around 90% to 65% at equivalent performance. Also a GTX1060.

So if we do math, you have almost archieved the same perf/watt as me, if it´s roughtly... :

*100% TDP : 120W
*90%TDP: 108W = Your power consumption stock
*65%TDP: 78W = Your power consumption with undervolt

I was monitoring all with afterburner before i discover that you can "Link" values from other programs like Hwinfo64 with Afterburner and have more values combined.

Of course in a daily basis, all of the monitoring apps are disabled, they´re so annoying; more or less like the Afterburner Skins

 

[Smanci]

How similar are the 100% TDP values set in VBIOS between 1060s?

With 970s there was a lot of variation - Gigabyte G1 100% TDP was 250W while ref only 145W.

 

[Raxe]

How similar are the 100% TDP values set in VBIOS between 1060s?

With 970s there was a lot of variation - Gigabyte G1 100% TDP was 250W while ref only 145W.

Certainly, i don´t know, but i suppose that with only 6 pin( my card atleast, i don´t know @Zeroth Alpha model) and only having slightly few Mhz vs gtx 1060 stock i don´t think that my card consumes a lot vs reference.

*Ok, you can see reviews that marks Gtx 1060 customs with 6 pin arround 125-130W peak, almost reference values like:

Msi Gtx 1060 6 pin: https://www.techpowerup.com/reviews/MSI/GTX_1060_OC/21.html

So, if you go to the 8 pin variants, there´s not a lot of discordance +20-25W more. Like:

Gigabyte Xtreme gtx 1060 8 pin: https://www.techpowerup.com/reviews/Gigabyte/GTX_1060_Xtreme_Gaming/24.html

Msi Gtx 1060 armor 8 pin: https://www.techpowerup.com/reviews/MSI/GTX_1060_Armor/28.html

They are arround 140-150w Peak.

You´re example in Gtx 970´s was so extreme, you cannot compare that! At least the Gtx 980 G1 has two 8 pin! It´s a massive sucker hahaha

 

[zovc]

With 970s there was a lot of variation - Gigabyte G1 100% TDP was 250W while ref only 145W.

I think the 250W TDP is them either indicating what they believe their cooler is rated for or what their maximum power delivery is rated for. The card's design and technology is going to be pretty much 1:1 between manufacturers, their custom boards will (as far as I know) normally have small tweaks that are primarily focused around their coolers and their desire to enable the card to draw more power for overclocking.

 

[Thehack]

I think the 250W TDP is them either indicating what they believe their cooler is rated for or what their maximum power delivery is rated for. The card's design and technology is going to be pretty much 1:1 between manufacturers, their custom boards will (as far as I know) normally have small tweaks that are primarily focused around their coolers and their desire to enable the card to draw more power for overclocking.

This is true. Funnily, they can also tweak the card for lower power consumption, by dialing in the lowest voltage and performance per watt, through the BIOS control. But no one does that because no one cares about power consumption enough.

 

[zovc]

Expect a pretty long post tonight about me playing with voltages on my 1070 tonight. Currently, I'm testing the max stable core clock in Unigine Heaven at each voltage. I might only have 'overclocking' results tonight, but I'm using this information as a baseline for my underclocking/undervolting and performance comparisons. How much less power and heat do you make when you lower your clock speeds at a given voltage? Is there a different sweet spot for each amount of voltage? Stuff like that.

 

[zovc]

We've hit a bit of a speed bump. My GPU doesn't seem to want to let me lock its voltage above 0.950V, and it's thrown a wrench into how I'm tuning and testing things.

I made a post on the overclocking reddit, I don't know where else to look for support.

I can share the google spreadsheet I have with my numbers so far, they're not nearly as complete as I'd like. Is there a 'private' way to share a google sheet that won't show off my email address and whatnot?

 

[Biowarejak]

I think if you just set it to "anyone with a link can view" it should be fine. Always did that in school. But I'm not sure if your email will be visible. You could also just screenshot it.

 

[Raxe]

We've hit a bit of a speed bump. My GPU doesn't seem to want to let me lock its voltage above 0.950V, and it's thrown a wrench into how I'm tuning and testing things.

I made a post on the overclocking reddit, I don't know where else to look for support.

I can share the google spreadsheet I have with my numbers so far, they're not nearly as complete as I'd like. Is there a 'private' way to share a google sheet that won't show off my email address and whatnot?

If you see my past tests if i go from 1911Mhz to 2000Mhz and from +350 to +400Mhz i lose some performance, 0.2fps; as same as you a little %(although it has more oc) so, my theory:

* Maybe the gpu it´s at his limit of frequency/voltage and there are some slowdowns making clocks a little bit lower o lowering voltages, idk, like a throtling
* Maybe cause they´re so little differences in frequency, they are little derivations, in your case is only 1 fps in a range of 190. it´s that like... -0.5% lower performance?

My sweet spot i have dropped it to 1810Mhz instead of 1835Mhz, cause strangely i win some performance at 0.85V.

Remember, as far i know, Gpu Boost 3.0 is not disabled vía this method only locked so, if you´re temperatures increase, the frequencies are gonna descend in -12Mhz or so, each 10 degrees starting a fairly low temp, 46 or 56 ºC...

 

[zovc]

My sweet spot i have dropped it to 1810Mhz instead of 1835Mhz, cause strangely i win some performance at 0.85V.

Remember, as far i know, Gpu Boost 3.0 is not disabled vía this method only locked so, if you´re temperatures increase, the frequencies are gonna descend in -12Mhz or so, each 10 degrees starting a fairly low temp, 46 or 56 ºC...

The only explanation I can think of is that it's GPU Boost doing its thing like you say, but the thing is, it's causing my score to go measurably DOWN from when I have the voltage and clock speed locked on lower settings. I've been trying to keep the card's temperatures stable, and it got to the point where I was running the fans at 100% (keeping the card around ~60C) during most of the ~0.900V+ testing. Prior, I had the fan set to my curve that's pretty lenient on temperatures in favor of quiet fans, but as the temps started getting towards 70C, I'd increase and lock the fans to try to keep the card in the 55-65C range. It probably doesn't help that Heaven never stops running so the card never gets a break, but it's also not like I'm running it at ~80C+.

Pretty consistently, as I have increased the voltage from 0.825V, my 'efficiency' has gone down. I need to do some better comparisons of numbers and I should be using a delta rather than the comparisons I initially came up with, but I was just trying to get data rather than draw conclusions at this point.

Again, right now these numbers are focused on finding the highest stable clock speed at each given voltage, not me trying to find the most efficient clock speed at a given voltage.

Things to note:

• The Watt numbers I give (Max and Average) during the benchmark aren't exact, they're rounded to the nearest Watt (up from .5), and the Average does have about ten seconds to one minute of non-benchmark time included. Nevertheless, Heaven keeps running even after the benchmark finishes so that shouldn't be too big of a deal. These Watt numbers were recorded via HWiNFO64.
• The Heaven FPS readings are rounded, nearly all of them were in the .0-.1 difference, but I believe there was one or two at .7-.9, either way I rounded up from .5 or higher. In hindsight I should have included the decimal place.
• The "Clock Speed/X" section is the row's Clock Speed divided by the value listed in the column. Voltage, Max W, Average W, in that order.
• The "Heaven Score/X" is similar to the clock speed section, but instead it's the Heaven Score divided by those values.

Here's the Spreadsheet I was working on.