Tesla k20 performance?

Hi, nice to see the forums back and looking all lovely! :)

We are thinking of purchasing a new card for development after our Tesla C2050 died a while back and I have been using a GeForce 580/680.

Has anybody tested a K20 or K20x compared to a GeForce 580 in relation to performance with OptiX ray-tracing?

My manager is very keen to go out and buy one of these cards but I would appreciate some information on the performance we may experience over the GeForce 580/680. The extra memory would be pretty handy though and it does sound like dynamic parallelism could be something OptiX may support now/later to great benefit!


Hi Craig,

results of K20 performance tests are mixed, but many use cases show a good speedup.
In a new version of DesignGarage we have seen 145 percent better performance compared to a GTX 580.

My manager went out and bought a tesla K20 for me which arrived a few days ago. I have been testing lots of combinations of cards to see what we get including a newer pcie3 and an older pcie2 machine. The results aren’t too great sadly.

Below are some fo the results for a test scene consisting of multiple views in a static scene containing 150 lights in an acceleration structure rendering @ 1920x1200:

GPU | Max | Mean | Min

580 13.571 9.883 7.065
680 17.021 11.962 8.569
680 (580 video) 16.485 11.830 8.409 (Relative to card on its own: 1.0325, 1.01116, 1.0190)
K20 (580 video) 17.455 12.465 9.060
K20 theoretical 18.022 12.604 9.232 (Based relative performance of 680 when using 580 for output)

K20 vs 580 132.8 127.5 130.7
K20 vs 680 105.9 105.4 107.7

As you can see from the k20 vs 680 results there ends up being little to no benefit (for what we are doing at present) for using a K20 in ray-tracing. We would expect that possibly in the correct cases dynamic parallism may be used (maybe?) for the ray-tracer or when the test scenario gets more complex there could be many more benefits.

I should also mention on the PCIe2 setup we saw a ~140% relative performance. This is no where near the 145% better performance stated in the Design garage but I am not sure if you mean the figure is relative and infact 45% better performance?


p.s. As a side note anybody needing more memory should definitly take a look at these cards or a Quadro as they are immense in that sense!

EDIT: Appears tabs/spacing doesn’t get kept by the forums :(

For some applications we see similar results. The K20 performs similarly to a 680 or a single GPU of a 690.

We hope to take advantage of new algorithms that are possible on the K20 in future versions of OptiX (dynamic parallelism), but we obviously can’t commit to specific features or timelines. ;)

Thanks James,

That reply puts our mind at ease as the dynamic parallelism sounds to have potential indeed. I only tested the one case sadly but will maybe get around to trying some more varied scenes to see what we get in the next few months.