-
CPU vs GPU and rendering speed
I’m running some speed tests with Arri raw files and am trying to analyze the results. Here are the approximate FPS (looking at the GPU meter while rendering) and CPU usage (looking at Activity Monitor while rendering).
MXF – 12fps, ~380% CPU
ProRes HQ 422 – 23fps, 450% CPU
10-bit 422 Uncompressed Quicktime – 34fps, ~325% CPU
10-bit DPX is hard drive-limited to around 30fpsIs this a GPU or CPU limited situation? This is a single-node one-light correction that plays back with solid green GPU at 24fps on the timeline. I’d assumed that the compression would be slower because it was CPU bound but I’m only seeing a minor increase in CPU activity vs. uncompressed. Can I speed MXF and ProRes rendering up with a PCIe expander and more GPU power? This machine has only the stock 6GB of RAM at the moment, could that be a factor as well? This is a 2010 2.4Ghz 8-core with GT120/GTX285.
-
7 Replies
-
Rendering, as far as I’m aware, is a CPU only process. You’ll see improvements where possible with the latest 12-core machines
Jamie Allan
Post Production Consultant
DaVinci Specialist (Linux/Mac)
Jamie@Jigsaw24.comJigsaw Systems Ltd. – IT & Broadcast specialists for the UK
https://www.jigsaw24.com
https://www.jigsawbroadcast.com -
Hi, in Resolve we calculate all grading processing, including for rendering, in GPU and with 8.0 due for release in June there are more debayer and other calculations also in GPU. The CPU’s are used for the encoding and decoding processes, like converting ProRes to RGB pixel data for the GPU’s. The CPU’s also are utilized for the disk I/O process just like any other disk I/O.
If your study is related to rendering, there are other considerations. For example, can the drive subsystem provide bandwidth to read the files on disk, send then to Resolve, have them processed and then write them back to the disk. Often its the Raid controller or disk I/O that’s the speed bump.
With two discreet volumes you can discount the disks themselves but unless you have a fibreChannel drive interface, and thus a separate Raid controller for each drive chassis, the Raid controller is one place to consider.
We have customers who can render at +100fps but it requires the right combination of hardware.
Peter -
The tests were done reading from a fiber channel array, rendering to an internal 3-drive RAID-0 array (3x1TB Western Digital RE3). The DPX test renders at 30fps so I don’t think the drive systems are limiting the compressed results. Is the GPU meter an accurate reflection of the amount of GPU activity or is it just a display of the current overall FPS? MXF rendering is slower than ProRes, yet uses less CPU as well. Does that mean MXF rendering uses more GPU for rendering than ProRes and would benefit from a second GPU?
-
What MXF codec are you rendering to?
-
Sorry, I guess that would help. It was DNxHD 1080p 220/185/175 10-bit. Project is 1920×1080/23.98 from Arri raw.
-
Peter,
Can you describe a modest Resolve system that will export 24 fps, 1920 x 1080, ProRes422HQ from Red 4K footage at better than 24fps?
As you mentioned above, you commonly see systems capable of 100fps+ renders. This blows my mind. How is it possible. And is there a happy medium?
Thank you in advance.
Chris O.
Chris M. Oben
-
With Red footage, you are limited by the Red decode speeds, typically Red Rocket which does about 22-23 fps for Epic 5K. So not matter how many GPUs you use or how fast your CPUs are, you are limited the source codec.
Other formats like ProRes/DNxHD/CinemaDNG/ARRIRAW, etc. can go a lot faster as they are not limited by the source codec.
So if you need faster Red speeds, you can consider multiple Red Rocket cards!
Reply to this Discussion! Login or Sign Up