Render farm - client nvenc and main none, can this work?
I've been using cinelerra for years with software encoding. The size of my projects has never been big enough for this to be a problem (just render over lunch break). My machine is an ultrabook with a 1.8GHz i7 8th gen quad-core chip (so 8 threads) and a dinky MX150 nvidia card (no nvenc). Recently however the render times have been becoming a problem. Also, I now have a second machine (desktop) with an AMD Ryzen 5 3600 and an nVidia 1660. I've set up renderfarming and it seems to work fine (my one and a half hour render is down to half an hour). Now I'm wondering whether that nice graphics card can be used (without me having to actually sit at the desktop, I'd much rather work/edit on my ultrabook with all its attached monitors) for editing. It seems whether nvenc gets used depends on whether I select h264-nvenc (or h265-nvenc) on the main machine, but obviously I can't do that on this particular main machine. Is there a way to get this to work, or am I forced to choose between software renderfarm and single-machine nvenc (in my specific situation, I'm assuming that if I had two nvenc-capable machines this would work fine). Or is there some cuda-related speed up I could use?
Discussed this with GG and he provided the following tips based on a video that plays in 1 hour running at full speed. In his final render on an AMD desktop computer with many threads, he breaks this up into 20 threads using a local render farm (just the 1 computer) so it only takes about 5 minutes to render the whole thing -- in your case on your Ryzen desktop with 16 cores a smaller number would be good. such as 16 maybe. One thing he said to be sure to do to optimize this is to "Reset rates" in the Render Farm menu so that the jobs are evenly distributed. In deciding how many threads to use, he suggests looking at the output of the linux command "top" to see how much cpu is being actually utilized on a typical render. Recently however the render times have been becoming a problem. Also, I now
have a second machine (desktop) with an AMD Ryzen 5 3600 and an nVidia 1660. I've set up renderfarming and it seems to work fine (my one and a half hour render is down to half an hour). Now I'm wondering whether that nice graphics card can be used (without me having to actually sit at the desktop, I'd much rather work/edit on my ultrabook with all its attached monitors) for editing.
I am not quite sure if I understand this correctly, but I would do all of your editing on your laptop first and when time to render just move the project to the desktop and do the final render there using the render farm.
It seems whether nvenc gets used depends on whether I select h264-nvenc (or h265-nvenc) on the main machine, but obviously I can't do that on this particular main machine.
Is there a way to get this to work, or am I forced to choose between software renderfarm and single-machine nvenc (in my specific situation, I'm assuming that if I had two nvenc-capable machines this would work fine). Or is there some cuda-related speed up I could use?
Although I can not test whether a software render farm or a double-machine nvenc is faster, gg thinks that a localhost render farm would be faster than using 2 nvenc graphics card enabled computers because there would be contention on the path to the graphics cards. It could be slower to have 8 threads using 2 cards than having 16 software threads. There does not appear to be a cuda-related speed up available as far as I know. Perhaps, I have confused this answer. If so, please reply.
Oh that sort of speedup would be very good, let me play around with reset rates. Currently I'm using 50% CPU on the desktop and about 60% on the laptop when rendering, definitely room for improvement. I will also compare with desktop-only (nvenc), but if I can double the current speed I get I'd be more than happy. Thank you very much! On Fri, Oct 23, 2020 at 5:48 AM Phyllis Smith via Cin < [email protected]> wrote:
Discussed this with GG and he provided the following tips based on a video that plays in 1 hour running at full speed. In his final render on an AMD desktop computer with many threads, he breaks this up into 20 threads using a local render farm (just the 1 computer) so it only takes about 5 minutes to render the whole thing -- in your case on your Ryzen desktop with 16 cores a smaller number would be good. such as 16 maybe. One thing he said to be sure to do to optimize this is to "Reset rates" in the Render Farm menu so that the jobs are evenly distributed. In deciding how many threads to use, he suggests looking at the output of the linux command "top" to see how much cpu is being actually utilized on a typical render.
Recently however the render times have been becoming a problem. Also, I
now have a second machine (desktop) with an AMD Ryzen 5 3600 and an nVidia 1660. I've set up renderfarming and it seems to work fine (my one and a half hour render is down to half an hour). Now I'm wondering whether that nice graphics card can be used (without me having to actually sit at the desktop, I'd much rather work/edit on my ultrabook with all its attached monitors) for editing.
I am not quite sure if I understand this correctly, but I would do all of your editing on your laptop first and when time to render just move the project to the desktop and do the final render there using the render farm.
It seems whether nvenc gets used depends on whether I select h264-nvenc (or h265-nvenc) on the main machine, but obviously I can't do that on this particular main machine.
Is there a way to get this to work, or am I forced to choose between software renderfarm and single-machine nvenc (in my specific situation, I'm assuming that if I had two nvenc-capable machines this would work fine). Or is there some cuda-related speed up I could use?
Although I can not test whether a software render farm or a double-machine nvenc is faster, gg thinks that a localhost render farm would be faster than using 2 nvenc graphics card enabled computers because there would be contention on the path to the graphics cards. It could be slower to have 8 threads using 2 cards than having 16 software threads. There does not appear to be a cuda-related speed up available as far as I know.
Perhaps, I have confused this answer. If so, please reply.
-- Cin mailing list [email protected] https://lists.cinelerra-gg.org/mailman/listinfo/cin
There has already been a long discussion. See if it can come in handy: https://lists.cinelerra-gg.org/pipermail/cin/2020-June/002116.html
participants (3)
-
Andrea paz -
Oon-Ee Ng -
Phyllis Smith