Blame Arnon

Nvidia GTX 1080 vs Nvidia Tesla K40 – Installing a GTX1080 in a Dell R720

At SQream Technologies, we use Nvidia graphics cards in order to perform a lot of the heavy database operations.
With SQream DB, we usually recommend using a Tesla K40 or K80 card. While a Tesla K40 is designed to operate inside a server enclosure (it has no onboard fan), standard Nvidia cards like the GTX series are designed to be run inside a regular chassis, and have an onboard fan to cool things down instead of relying on the server fans.

Sometimes, customers have different demands and don’t want to pay the premium for the Tesla cards that we prefer. We therefore wanted to check if we can offer our solution based on the GTX 1080.


We received a loaner GTX 1080 in early June for our tests. We decided to benchmark the GTX 1080 against our incumbent Tesla K40 (which is very similar to half a K80 in most situations), in a server environment. For that task, we decided to run the TPCH performance suite on SQream DB.


We recompiled our database with the CUDA 8 release candidate and inserted the GTX 1080 inside the Dell R720 server enclosure.
Snug GTX 1080 in a Dell R720


Our results show that the GTX 1080, despite being much more powerful, was only 13% faster than the Tesla K40 at most operations, including the relatively complex TPCH-5 query.

Simpler queries like TPCH-1 did not benefit as much from the GTX 1080, and only showed a 3% performance increase.

The Tesla K40 is much more expensive however, and uses almost twice the amount of power. It averaged around 150 watts during intensive load, compared to around 80 watts with the GTX 1080.

Our main concern was overheating. We were happy to see that the GTX 1080 runs nice and cool, at around 40-50 degrees under intensive load inside a Dell R720 enclosure with fans on maximum output.


While we were pleased to see that the GTX 1080 was not slower than the K40, but it wasn’t fast enough to justify a move to this card for most users, with a 3%-10% speed improvement for intense SQL queries.
We were also very pleasantly surprised to see that the card did not overheat and used significantly less power when compared to the high powered Tesla series.

Leave a Reply

  • jesus says:

    Arnon, What about the performance of new gtx 1080 for double precision? Did you compare both cards?

    • Arnon Shimoni says:

      We didn’t test DP specifically, but it was included in our larger suites of tests. Performance is roughly 15%-40% better with no code changes

  • Guanji says:

    Hi Arnon,
    I have a GTX 1080 GPU card and want to install it on a R730 server. I have prepared the GPU enablement kit and get the card with external power cable connected, but I still get the ‘Unable to communicate with GPU because it is insufficiently powered’ error. How do you have your GPU powered up in your R720 server?

    • Arnon Shimoni says:

      Hey Guanji,
      Did you connect the cable including all pins?

      The GTX1080 has an 8-pin connector, and all pins must be connected from the riser card to the GPU.

  • Tang Thomas says:

    Hi Arnon,

    Thanks for sharing!

    Your article inspire me to install GTX1080Ti into my R720 with Ubuntu 16.04, CUDA 8.0 & nvidia-384.

    I can see the device there by nvidia-smi, and I pass the sample run for deviceQuery, bandwidthTest.

    But, I can’t get GPU in nvidia-settings, nor pass any graphic sample run.

    What’s the version of OS, CUDA & Driver using in your R720 + 1080? I’m tring and stucking there for a while.

    It will be great, if you can share your installation & verification process in next article.