Agisoft Metashape

Agisoft Metashape => General => Topic started by: an198317 on February 07, 2014, 09:06:35 PM

Title: Nvidia Tesla K20 Graphic card vs Geforce Series
Post by: an198317 on February 07, 2014, 09:06:35 PM
Hi everyone,

We are going to put together a workstation for Agisoft PhotoScan Pro processing. We were looking at Nvidia Tesla K20 card. But I found a 3-yr old thread talking about this very topic. But that was about Tesla C2075.

And I also found Alexey's response as following:
"PhotoScan uses single precision arithmetics in the GPU processing part (actually the most time consuming calculations are done even using integer arithmetics). So the double precision performance limitations of the consumer video cards doesn't affect the processing speed at all. Consumer GeForce video cards should be as fast as professional (Quadro or Tesla) ones for PhotoScan."

Question 1: Does the new Photoscan Pro 1.0 version still use single precision arithmetic for the GPU processing?

Question 2: I can't find any Geforece GTX series specs on peak single precision floating point performance. But the Tesla C2075 is 1.03 Tflops, and the K20 is 3.52 Tflops, and K20 is significantly faster than C2075. So will K20 be faster than consumer level Geforce GTX series?

Thanks!
Nan
Title: Re: Nvidia Tesla K20 Graphic card vs Geforce Series
Post by: Alexey Pasumansky on February 10, 2014, 01:57:59 PM
Hello Nan,

Yes, only single precision calculations are used.

By our assumptions Tesla C2075 / K20 will show comparable performance with GTX 780, but I should not expect that they will be faster.