Nvidia Pascal and AMD Polaris, starting with GTX 1080/1070, and the AMD 480

I'm gathering more data, and need to think some more, but my preliminary conclusion based on various measurements and subtractions (all system wall power based), is that adding two GPU tasks on a stock clock 1050 adds 54 watts to system power consumption.  I estimate a little over 3 watts of that goes to the CPU and motherboard for support, leaving just about 50 watts of the delta going into the 1050 itself. 

A persistent unknown is the card idle power.  I can say that in my system box idle power is down by over 3 watts when I have the 1050 in as compared to the 750.  This is actually quite nice, as I think the 750 is not bad for idle power (though not so good as it self-reports).  My working guess is that my 1050 idle power is 2 watts.  If true, then at stock clock running 2X BRP4G on this system the 1050 card consumption proper is in the low fifties of watts.

I'm consuming more idle power than necessary, as my fan curve bottoms out at 40%.  Sometime soon I plan to exercise the fan across full range in a quiet room with a view to assessing subjective tone and other undesirable behavior.  My preliminary assessment is that the particular card I chose (Gigabyte GV-N1050OC-2GD) has low fan noise of pleasing character in my use conditions so far.  Running it overclocked in a two-card case with less than excellent case ventilation will be more challenging, but I am currently optimistic.  People with less sensitive ears, better ventilated cases, and no intention to run two cards in a case can probably pay little attention to the fan issues on 1050/1050Ti card.

For overclocking trials I am currently specifying offsets from default using MSIAfterburner.  I started with +400 to Memclock together with +100 for core clock.  I soon hit a ceiling in which the card did not report advance to a higher memory clock when I specified +550 instead of +500.  It appears that there may be a hard limit to memory clock, at a speed represented by GPU-Z as 2002, by MSIAfterburner as 4004, and presumably called 8008 by Nvidia marketing.

Actually +500 is not bad, but I'd like to have been able to explore up to actual failure, in order to set a speed enough lower to feel comfortable.  For the moment I am leaving the memory offset request at +500 and inching up the core clock request (have seen success at +100).

Both my 1070 and my 1060 run DDR5 Samsung memory at rates appreciably faster than 2002/4004, and the cards allowed them to run enough faster still to get failure.  But I think I've seen reports in reviews of hard limits on the 1050 series cards.  A last point on memory is that GPU-Z reports this card to be populated with Hynix RAM, so neither the Micron nor the Samsung I've previously heard reports of on Pascal cards.

My 1050 overclocking work seems to have reached ceilings on both memory and core clock imposed by the card, firmware, or driver, without getting to an error of result condition.

The core clock actual available increments seem to be something near, but not exactly 12.5.  My last two offset requests which seemed successful were for +225 and +250.  The +275 request, which should have gotten a result two "ticks" higher got the same speed.

At current operating conditions (temperature matters) this means I'm running at an indicated 1911 core clock, 2002 memory clock on the GPU-Z reporting scale.  I plan to run overnight at this condition to get an initial safety assessment and collect operating condition (power, temperature) averages.  All this has been at 2X on BRP4G/Cuda55 on Window 7 with the latest Nvidia driver as of a couple of days ago.  I intend to run brief trials comparing 1X, 3X, 4X, 5X.  

While at one level it is annoying not to be allowed to run the card as fast as it can give correct answers, at another level this might be viewed as a sort of safety cushion which helps make this a particularly suitable naive or entry level user card, unlikely to hurt the project by being set to run a little too fast by an inattentive user who returns flocks of invalid results.