I thought I would start compiling a list of how quickly various GPUs finished the new OpenPandemics GPU workunits (OPNG), and some baseline comparisons between various kinds of cards. I also thought I would caution against using online databases to compare cards for the purposes of OPNG. While I only have about 160 data points (so these are very much preliminary figures), I thought we could start cataloging to see if there are any cards that are a particularly good value.

Baseline assumptions: All of my cards have been computing with stock parameters. No overclocking, 1 CPU assigned per OPNG workunit, 1 workunit assigned per GPU. That is to say, I have not yet implemented any changes to <gpu_versions> in my app_config.xml or cc_config.xml files, and I am assuming wingmen also have not. I am also assuming the time reported in the workunit is an accurate measure of how much time the GPU was crunching the unit. Finally, I assume that the computation time linearly increases the more "jobs" each OPNG unit has. So, if it takes a GPU 10 seconds on average to compute one "job", then a workunit that has 50 "jobs" should take 500 seconds.

That said:
I have a Quadro M2000, a GT 1030, a GTX 1650, and an RTX 3080. The average time per "job" of each of these cards is:
- M2000: 13.04 seconds / job (50 data points)
- GT 1030: 15.84 seconds / job (14 data points)
- GTX 1650: 7.99 seconds / job (19 data points)
- RTX 3080: 2.45 seconds / job (15 data points)

I found that a number of OPNG units that were crunched on one of my cards had multiple wingmen with a singular card model that also crunched that same unit. While I can't control for RAM, CPU, Storage, etc, it may still allow us to see how cards stack up against each other. If the things that I can't control for don't majorly affect the numbers, then I can conclude from these comparisons:
-For every job that a Quadro M2000 crunches, a GTX 1070 can crunch 1.48 jobs (10 comps).
-For every job that a Quadro M2000 crunches, a GTX 1080ti can crunch 1.9 jobs (5 comps).
-For every job that a GT 1030 crunches, a 2070 Super can crunch 5.79 jobs (9 comps).
-For every job that a GTX 1650 crunches, a 2070 Super can crunch 2.4 jobs (4 comps).
-For every job a GTX 1050ti crunches, a RTX 3080 can crunch 4.66 jobs (2 comps).
-For every job a GTX 970 crunches, an RTX 3080 can crunch 3.01 jobs (2 comps).
-For every job a GTX 1650 crunches, an RTX 3080 can crunch 3.17 jobs (4 comps).
-For every job a Quadro K620 crunches, an RTX 3080 can crunch 15.31 jobs (4 comps).

Finally, when comparing cards using a site like TechPowerUp, I would caution against taking their rankings as gospel. Sure, general trends will hold - older cards will be slower and newer cards will be faster. But older cards may not be as slow or fast as those rankings imply. For example, TechPowerUp lists the GT1030 as being around 7% as powerful as an RTX 3080. With a 3080 taking an average of 2.45 seconds/job, TechPowerUp would predict that a GT1030 would take between 30-35 seconds/job. But with my 14 data points, a GT 1030 only takes an average of 15.83 seconds/job. And I have direct comparisons between a GT 1030 and a 2070 Super where the 2070 Super takes about 2.8 seconds/job (which is the overall average I've observed) and the GT 1030 takes 14.89 seconds/job.

I am also surprised at how well 1660s, 1650 supers, 1660 supers and 1660tis fare compared to powerful, older cards. They all seem to outperform GTX 1080s and 1080tis. The 1080ti has an average of around 6.2 seconds/job, but the newer cards seem to have an average time / job of around 4.3 - 5 seconds/job. And speaking of newer cards, they all seem to be closely grouped, despite TechPowerUp showing some significant drop-off in power. For example:
3080: 2.45 seconds/job (15 data points)
3070: 2.65 seconds/job (1 data point, ~9% slower instead of 23%)
2080 Super: 2.69 seconds/job (2 data points, ~9% slower, instead of 31%)
3060 ti: 2.79 seconds/job (1 data point, ~10% slower, instead of 30%)
2070 Super: 2.88 seconds/job (15 data points, ~15% slower instead of 39%)

What about you guys? Have you noticed any patterns? Have you observed different numbers? What do we think makes the most difference in speeding up the seconds/job? Memory bandwidth? CUDA cores / Stream Processors? Frequency? Next, I intend to start testing how modifying the <gpu_versions> affects these numbers.