GT 1030 ............ Pascal GP108-300-A1 ................... 1228 MHz ..... 14 nm Samsung
GTX 750 Ti ....... Maxwell GM107-400-A2 ................ 1020 MHz ..... 28 nm TSMC
GT 730 .............. Kepler 2.0 GK208B ........................... 902 MHz ..... 28 nm TSMC
GT 710 .............. Kepler 2.0 GK208 .............................. 954 MHz ..... 28 nm TSMC
GeForce 210 ... Tesla 2.0 GT218-300-A2/ -325-B1 ... 520 MHz ..... 40 nm TSMC
G550 ................. Condor ................................................. 125 MHz ... 180 nm UMC
These data are not really indicative of idle power usage though.
The Radeon 6450 arrived today. It's a very basic passively cooled card with a fairly small heatsink. I got my dual boot setup and switched over to Linux for the time being to run some Wanless(but with the Matrox the power usage was identical between running Numberfields in Windows and Wanless in Linux). The power usage is 130-131 watts currently with the 6450. If I start really moving Windows around, trying to make the GPU work a bit it will get up to 133-134 watts. The card cost $13 on eBay including shipping, and has an HDMI port which is nice(the Matrox is just 2 DVI ports). It seems like either of these is a good bet for my purpose. The 750ti power usage was good enough, and it could potentially be used for some low PPD folding or something(increasing the power usage), but it was also much more expensive than the other cards. For my purposes, either the 6450 or the Matrox both seem good, so each will go in a machine here.
I think I'm going to switch to 65 watts eco mode and compare the task times on Wanless to PPT being set manually to 88 watts currently.. I'm guessing where it really would take a huge hit would be on AVX2 workloads, though.(maybe?). Might be worth running some Primegrid as well. I'm not sure how low I'd have to get the power usage to be comparable to EPYC. 130 watts from the wall for 24 fairly highly clocked threads didn't seem too bad to me, though.
Edit - at 45 watt eco mode it's using 94 watts from the wall. Considering all that is in it is the 6450, 2 sticks of DDR4 and an NVME solid state drive, I'm surprised it's not a bit lower. The power supply is probably something around 85% efficient at this power usage(the efficiency curve I found starts at 20% usage, this is around 15%, so I'm guessing the efficiency is a bit lower than at 20%). That would mean outside of the 45 watts from the processor, taking the power supply efficiency into account there is 35 watts used by the 6450, 2 sticks of RAM, an NVME SSD, and 2 fans and then the chipset and everything on the motherboard. The Wanless tasks haven't completed yet so I'm not sure how much less PPD it is at 45 watts eco mode yet.
Edit 2 - using 45 watt eco mode was pretty efficient. It was completing Wanless tasks in 2800 seconds set that way, and 2400 seconds with PPT set manually to 88 watts. Actual power usage in 45 watt eco mode(from the wall) was 94 watts vs 130. I'm trying 65 watt eco mode now, but interestingly enough it's only 1-2 watts different than what I was getting at 88 watts PPT set manually. It's at 129 watts currently set on 65 watt eco mode.
As you probably know, I preferred to build dual-socket systems. Benefit: Half the amount of mainboards, PSUs, disks, network interfaces, BMCs/GPUs (and potentially: cases) compared to single socket systems. Also, half the amount of operating system instances to set up and maintain.
Half the amount of disks and NICs etc. also means better perf/W at the system level. But there is a downside in the perf/W department: These dual socket systems maintain a high bandwidth/ low latency link between the two sockets, which eats energy. Now, in virtually all distributed computing applications, actual traffic on this link is very low in practice. Whether or not this keeps energy use of the link low is something which I haven't found out yet.
I could see picking up a 1010 to replace my 710 I have sitting sitting around as a spare. The old 710 has gotten used more than I expected, and the 1010 is certain to be faster and hopefully more efficient.
I presume though that GP108, like the larger Pascal chips, does not include a D/A converter for VGA support anymore. GT 1030 cards with VGA output exist, but they are rare and I therefore assume that such cards have an extra D/A converter onboard. Which is in contrast to GT 710's Kepler which natively supports VGA.