Tesla k80 vs. GTX 1080
Hi DSPH developers,
Do you have some experience using both the GPU card Tesla k80 and GTX 1080? I found that the computing efficiency of GTX 1080 is higher than Tesla k80 when calculating the same case (about 300,000 particles). In theory, the computing efficiency of k80 should be higher than GTX 1080, since k80 has 4992 cores, while GTX 1080 only has 2560 cores, and the price of k80 is much more expensive than GTX 1080. Is that right? Or can you give me some directions (such as flag in the compile command) to increase the efficiency of k80.
Regards,
Do you have some experience using both the GPU card Tesla k80 and GTX 1080? I found that the computing efficiency of GTX 1080 is higher than Tesla k80 when calculating the same case (about 300,000 particles). In theory, the computing efficiency of k80 should be higher than GTX 1080, since k80 has 4992 cores, while GTX 1080 only has 2560 cores, and the price of k80 is much more expensive than GTX 1080. Is that right? Or can you give me some directions (such as flag in the compile command) to increase the efficiency of k80.
Regards,
Comments
The number of cuda cores is only one side of the problem ! GPU cards have different clock speed (i.e. speed at which the card run).
K80, 4992 cores x 560 Mhz clock speed x 2 = 5591 GFlops in single precision
GTX1080, 2560 cores x 1607 Mhz clock speed x 2 = 8228 GFlops in single precision
CCL GTX1080 > K80, Yeah your results are normal.
you can check these data on the wiki page of nvidia product
https://en.wikipedia.org/wiki/Nvidia_Tesla
https://en.wikipedia.org/wiki/GeForce_10_series
DSPH code use double precision only when necessary (if I'm remenbering right what the devs told me), so using data on single and double precision computing and GPU prices you can make a reasonable choice of GPU to buy.
Also you should be aware that all the gamer card are not the same. Depending of the manufacturer, they are over-clocked at different speed (this is always specified so you can check it). This over-clocking can boost your performance a lot !! You also can over-clock the card yourself, but it is not trivial and you loose the guaranty, so...
I think "professional" cards are expensive and have lower speed clock, because they are more reliable than "gamers" cards. They have systems that ensure no errors are made in calculus and report of calculus. This has a cost in term of hardware price and running speed of the card. Not checking for errors could be desastrous for some applications (AI, neural networks...) but is not really a problem for SPH. If a small error happen for one particle among millions of them...
Hope that this can help you
Tanguy
Cuda cores are brains, their clock speed represent how fast they think !
regards
the Tesla K80, has 2x GK210 chips. So having a K80 is like having two graphic cards in your system. DualSPHysics will only use one of the cards so you are actually using half of the processing power available. So the number of usable cores is similar in both scenarios but the frequency of the 1080 is much higher.
Regards!