Some time ago we tested running 2 WUs concurrently on Kepler GPUs, like GTX660Ti. There a clear benefit was observed for short runs, whereas long run WUs lost performance. The reason was that GPU utilization was already good enough, so switching between WUs only introduced further overhead.

With my new GTX970 running the recent Gerard CXCL12 WUs at just 7x% GPU utilization I found an approximate throughput increase of 12.5% with 2 concurrent WUs. GPU utilization went up to 91% (the WDM performance tax of Win Vista and newer applies).

I've been running like this since then, because there are so many of these "small" WUs (in terms of number of atoms) are still being distributed. But some of the usual big Noelias also crept in. Now the interesting part: I'm also seeing a throughput increase for these :)

They used to take about 21.5 ks each, with 162 W power draw and ~85% GPU utilization. Now two of them finish in about 38.5 ks. This includes some time of each WU running solo, when the other one uploads. So let's assume 2 WUs every 39 ks.. which yields a 10% throughput increase! Note: since then I've increased the power limit, so that my card now consumes 168 W (3.7% more), but runs approximately at the same boost state.

I assume this also applies to similarly powerful cards: GTX780Ti, Titan, Titan Z, GTX970 and GTX980. It may also work for GTX780 and GTX770, but I would expect the benefit to be lower.

It could also be that some improvement in Maxwell results in more efficient task switching, but I have no further information to judge this.

MrS
____________
Scanning for our furry friends since Jan 2002

The amount of GPU memory is a known factor; if it doesn't have enough memory to run 2 tasks then it will be much slower overall.
As the 970 has 4GB GDDR5, memory amount isn't an issue for 2 tasks ATM.
My guess is that some other cards with 4GB might be capable of seeing improvements even around 85% GPU usage, but there is still going to be a point (say 90%) when you are not going to see any improvement, if not a performance loss.

I've never tested this with a Kepler GK110 and they are architecturally different from GK104/GK106/GK107 where I think there is less chance and you are really limited to the GTX660Ti 3GB models and up (670 4GB, 680 4GB, 770). Possibly the 760 4GB too but it's only got 1152shaders which maybe an obstacle in itself.

I expect that performance to vary by GPU model (Big Kepler's, small Kepler's, Maxwell's, bus width) how you tune it (GPU/RAM freq.), task type and the operating system (as GPU usage might be higher on XP and Linux to begin with).

I had a concern with the GTX660Ti 2GB that even with smaller molecules (lower GPU usage) some RAM was not always as easy to access (slower), beyond 1.5GB IIRC. So for 2 tasks of <750MB it might have been ok, but not >750MB. Might have been related to the super-scalar architecture, which Maxwell's don't have. Note than Win Vista and above uses 30MB or so on GPU0, so for that card the tasks would need to be <735MB.

The cache size might also come into play; the bigger the better.
GK104-300-KD-A2/GK106-400-A1/GK107 isn't as large (384K/384K/256K).
The other GK104 models (GTX770, GTX680, GTX670) have 512K.
GK110 is 1536K and GM107 is 2MB (though it's only got 1GB or 2GB GDDR5).

I expect the MCU load would be higher with 2 tasks, so relatively speaking, the 970 might be slightly better than the 980 as the 970 has a lower load/(shaders/bandwidth ratio).
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Yes, the GTX660Ti and Noelia WUs in general won't benefit from 2 concurrent WUs. And regarding the power limit: the percentage is not relative to the TDP, but rather to the power target. Which is "just" 130 W for normal GTX660Ti's, so it's not that bad.

My cards also run into power limits - but this is on purpose. By lowering the power limit (target) I'm keeping the 28 nm cards around 1.10 V, where the clock speed is almost as high as at full boost (loosing around 70 MHz), but the energy efficiency is ~20% better.

MrS
____________
Scanning for our furry friends since Jan 2002

Yes, GPU-Grid is pushing the GPUs quite hard, pretty much like demanding games. This is not to say Einstein woudln't use them efficiently, but their algorithm is "less arithmetically dense", which means the have to spend more time on moving data around rather than "just" crunching the numbers.

If you feel your GPUs are working too hard simply lower the power target / limit. the cards will adjust clock speeds and voltages accordingly and run more power efficient (due to the lower voltage). I'm aiming for ~1.10 V on 28 nm GPUs, which costs me 50 - 70 MHz compared to full throttle operation, but saves ~15% power consumption.

MrS
____________
Scanning for our furry friends since Jan 2002