My Results: Power Draw 290 xFire w/ undervolt/underclock <550w performance

[Attic]

After getting into GPU mining I became aware of how much power can be cut back by undervolting and underclocking GPU's. I put this to use in a constrained cabinet that houses my HTPC below my TV. The power draw difference on my 290's from stock to underclocked is significant, ~25%, while performance loss is closer to 10%.

*In testing I also saw that the PCS+ cards throttle at stock due to the +50mv they get by default if the power limit is not raised by the user. Throttling appeared due to power target not temps IMO.

My setup:
Corsair AX860 Platinum Rated PSU
-4690k stock (add 30-50 watts to each result to mimic 4.2-4.6 OC)
-(2) 290 - Powecolor PCS+ Stock 1040/1350 w/ +50mv and +0 Power Target *yes, these cards come factory overvolted by +50mv
-ASIC Quality is 79% and 80%
-6case fans/2 SSD's/Optical Drive/4x4gb RAM
-Graphics Driver: 14.7 RC1 *Highest quality AF set via driver and no optimisations, everything else default

Using a Kill A Watt P4460 Power usage monitor. Power draw measured from the wall.

System Idle Power at Desktop is ~60w.
-Thank you AMD Zero Core and Intel power saving features

Power Draw Results taken over the course of 3 Runs of Firestrike Extreme and 2 Loops of Valley Benchmark (Ultra/8xAA/1080P/Vsync Off). Firestrike Extreme scores given and power use numbers are my interpretation of avg as well as noting the peak power draw.


900/1250 -50mv, +0 Power Target
Peak Power: 510w
Avg Power: 460-500
Firestrike Extreme GPU Score: 8454, 8433, 8450


947/1250 -37mv, +0 Power Target
Peak Power: 556w
Avg Power: 480-545
Firestrike Extreme GPU Score:8823, 8822, 8834


947/1250 +0mv, +0 Power Target
Peak Power: 573w
Avg Power: 520-550
Firestrike Extreme GPU Score: 8811, 8812, 8815


1040/1350, +50mv, +0 Power Target
Peak Power: 689w
Avg Power: 620-675
Firestrike Extreme GPU Score: 9460, 9373, 9410

*GPU1 throttles 1000-1040 w/ temp peaking at 91c, GPU2 locks to 1040 w/ temp peaking at 81c.

I can't open up Power Target at OC'd settings due to heat concerns, but this would increase power draws, perhaps substantially.

Performance of course is less while undervloting/underclocking, but with some of the non reference 290's I think my results can be duplicated at 947/1250 (stock 290) and 900/1250 and can be had with substantial power/heat savings vs what popular reviews show. Needless to say, 2x 947/1250 is very, very quick. 900/1250 is as well and is where the real power savings came through at my testing.

The 1250 on the memory is an important constraint (it should be kept at 1250 min), for reasons i'm not completely clear on. The voltage affects core and memory AFAIK and dropping voltage can upset the memories ability to run 1250. While core might run at -75mv, the memory doesn't like anything lower than -50mv to be stable at 1250 IME. Going below 1250 on mem appears to have big impact on my cards performance, perhaps due to timings. Both cards use Elpida Memory.

 

[PPB]

Memory voltage should be different to vcore. What is clear is that, as shown by the stilt, upping the mem clocks past some milestones produce loosened timming, which affected hash rates.

Uc/uv is even more interesting than overclocking in my view. The push for efficiency should be better regarded, as there is alot to gain in that area.

 

[Attic]

Memory voltage should be different to vcore. What is clear is that, as shown by the stilt, upping the mem clocks past some milestones produce loosened timming, which affected hash rates.

Uc/uv is even more interesting than overclocking in my view. The push for efficiency should be better regarded, as there is alot to gain in that area.

Uc/uv? The efficiency aspect is interesting to me in areas where a 2-1 or greater savings of % watts vs % performance can be attained. The 290 through certain points gets ridiculously inneficient. Certainly in mining without much tweaking it was much less efficient than 7970/280x, but when both were tweaked to full the cards were similar in efficiency.

I'm using MSI AB for adjusting clocks and volts and only have one voltage adjustment on the Powercolor PCS+ 290's.

I think that AMD sacrificed a lot of power and heat by making the R9 290 a 947mhz stock part instead of 900mhz stock part. I'm not sure why they pushed it so hard on that crap cooler and sent out a part to reviewers that was power hungry/hot/loud. A 900/1250 part as reference that doesn't suffer those issues, but can overclock like crazy if you want to I think would have had better initial reception. Non reference could have done the same thing it did in any regard, but the stigma (somewhat well deserved) of the 290 Reference still is there.


In further testing I saw a ~20w reduction by keeping temps 78c and 68c for GPU 1 +2 vs 90c and 88c when running 900/1250 clocks in crossfire. In my setup I will sacrifice some efficiency for noise, if I let the cards run to ~82-84c I can get by with a lot less gpu fan noise than by keeping them sub 80c. But even here at 84/82 there is decent savings from 90c+temps that the parts were largely reviewed at.

 

[thilanliyan]

I'm getting a draw of ~600w (measured from my Corsair 1200i which can measure power draw on its own) from a 990FX system (AMD FX6100) running 3x290s at 950/1250 with -50mv/-25mv, power limit at +20%. This is while mining X11/X13/X15 coins. This is with a bunch of watercooling gear and many fans though so actual draw from the cards is less of course.

 

[Headfoot]

I can vouch that my XFX reference 290 with Elpida memory also runs like crap at anything other than 1250. Lower or higher makes it hash a lot slower, sometimes noticeable FPS difference sometimes not depending on the game. Can get it to run 1000/1250 @ -25mv

 

[Silverforce11]

Elpida on R290 is terrible compared to Hynix.

When I was mining, I ran it undervolted at stock clocks and it was using about 200W each card (which is very low for mining load), lower than normal due to very low temps. I noticed a sharp drop in power use when you go below 60C and vice versa, and another sharp bump when you go above 87C (stock HSF).

 

[Shmee]

Yes the elpida memory is very finicky, I remember when I was playing around with clocks on my powercolor 290, it would bench much worse with anything other than 1250 memory.

I may see if I can undervolt the powercolor, it is actually in crossfire with a 290 tri X I got off ebay (hynix) so lowering temps is always desirable.

 

[Attic]

Yes the elpida memory is very finicky, I remember when I was playing around with clocks on my powercolor 290, it would bench much worse with anything other than 1250 memory.

I may see if I can undervolt the powercolor, it is actually in crossfire with a 290 tri X I got off ebay (hynix) so lowering temps is always desirable.

If your Powercolor is the PCS+ then the 290 Tri-x is a nice compliment being that it gets good/great temps being in top position if you've got 3 slot spacing between GPU slots. I noticed substantial temp improvement by getting good airflow not from the side, but rather fairly direct parallel through the cards from front to rear. I have both side and parallel air flow FWIW and that works well.

One thing I noticed about undervolting on the hynix mem. It doesn't like to be undervolted in Windows using MSI AB at mem > or =1250. Specific there because it undervolted fine for mining at 1000m. But in Windows with MSI AB for example, I would do 950core, 1250mem and -25mv and the system wouldn't immediatly fuzz screen for an instant and then crash after I hit "apply". If I recall correctly I couldn't undervolt the 290x Tri-x at all if I wanted to maintain 1250m. I actually chalk this up to advantage elpida in my situation. 850c/1000m was fine at around -50mv and hynix seems to have very tight timings at 1000 which somewhat makes up for deficit in clock rate.

I'll update OP with 1000c/1250m -25mv results. That's another nice spot.