Topic: [CureCoin] PPD Observations: Higher-End Hardware: Worth It! (Read 952 times)

Holy smokes!
We're starting to accumulate some serious power. This coin is going places, folks!

Still haven't opened the 2nd one, but it is oh so tempting, I put it out of my line of sight xD

The one 780 that I am running is doing well, it's hitting somewhere in the ballpark of 115000PPD right now, not terrible, but ATi cards seem a bit better for the price (especially 7970s).

I ordered a few more wifi cards, once they arrive I should be able to (temporarily) point another 5 or 6 7950s at folding, hoping to break the 1,000,000PPD bounty.

I kept meaning to ask in the curecoin pool thread how your 780(s) were testing, always forgot. (did you open the 2nd one?) 

Thank you for taking the time to do all this and posting your findings. Much appreciated.

For anyone wondering about the appropriate hardware for Folding@Home to maximize their PPD, go as high-end with cards as you can. The way the bonus points system works, even if two 660s can fold just as fast as one 680, the 680 will get you significantly more points. Here's why:

The bonus points model is based on how quickly you return the work unit, following this exact formula: PPD = base_points * max(1, sqrt( k * deadline_length / elapsed_time)) where k = 0.75

The idea is that you get a work unit, say a project 7810. Each project has a different amount of computations, and a certain guaranteed credit. For a project 7810 workunit, the base credit is 3076. As long as you solve that workunit (and of course upload it!) in the alotted time, you are guaranteed that many points. However, as the formula above shows, you get significan extra credit for submitting early. As an example on a 780, that 7810 workunit would be worth around 14500 points, which is 4.71 times the base credit.

Given this logic, it'd be (in terms of PPD), more beneficial to finish one workunit fast, than to finish four of them slowly.

A Bit of math to back this up:      
Project: 7811 Using calculator: http://www.linuxforge.net/bonuscalc2.php

TPF: 1 min 20 sec <-----NVidia 780
TPF: 2 min 40 sec <-----NVidia 760 (estimate)

The 780 would get 11292.67 points for each job, while the 760 would get 7985.03 points for each job.
While the 780 and the 2x760 builds would each get approximately equal numbers of jobs done per 24hr period, the 2x760 build would be at a disadvantage, making less PPD per kWh (2x760 and 1x780 use approximately equal power, 2x760 will be slightly higher due to having to run two sets of fans and other components that aren't duplicated when moving to a 780, however the main power draw is the cores and memory anyhow).

The machines with PPD compared:

780: PPD: 121960.8
2x760: PPD: 86239.34

However, it's important to note that 2x760 costs $520, while one 780 would cost $660.

This compared, the 780 gets 184.798 points per dollar per day (PPDPD?), while the 760s get 165.84489 PPDPD. As well, the 760s take up two slots, and consume slightly more power.

Higher-end cards are a given until you reach the Titan, at which ~140,000 PPD @ $1,000 is clearly not comprable to the 780.

You can see a similar relationship between ATi cards, a 7970 gets around 120000PPD, a 7950 hits around 80-90,000 PPD, and a 7850 gets in the 35,000 ballpark.

Of course, if you only want to buy a cheap GPU and fold a bit, there's nothing wrong with the lower-density, cheaper cards, which are great for entry-level builds. However, even if two 760s is the same power for less price, due to the bonus strategy, the 780 will produce higher yield PPDPD, at less electricity, with more room for additional components. As well, overclocks are very helpful, earning more of those sweet, sweet bonus points.