Login

Again on CUDA performances

Tips & tricks, working results, technical support

Again on CUDA performances

Postby Vernon » Wed Jun 08, 2011 3:39 pm

Hi,
I'm experiencing a poor performance using CUDA.
This is my configuration:
2 pc LAN connected (1Gbit)
Master pc: Intel Core 2 QUAD Q6600 2.4 GHz, 4GB RAM
Slave pc: Intel Core I5-661 3,33 GHz, 4GB RAM
OS: Windows 7 Ultimate 64bit
Sequencer: Reaper 3.76 64bit, using Reamote function to share plugins between pc's
Graphic Card: Nvidia GeForce GTX 570 (480 Cuda Cores) on master pc
Drivers: Nvidia 275.33, CUDA Toolkit 4.0.17, CUDA bridge v 1.1.038 (all x64)
Nebula3 Server WIN x64 1.3.492

I'm working on a large 44.1/24 mix (100 tracks), resources are 50/60% on each pc, none Nebula still loaded.
When I try to load a few instances (3/4) of Nebula Reverb (using for example the Alexb CLC Library buss group programs), pops and glitches starts and it's impossible to continue mixing.
I've tried previous CUDA toolkit/bridge (4.0.12/1.1.037; 3.0/1.036) with the same results.

I've tried, in order of performances:
- Nebula3 local server + CUDA
- Nebula3 local server w/o CUDA
- Nebula3 w/o server and w/o CUDA

So, the only way to load Nebula instances and work on the mix is to use the "stand-alone" version, without server or CUDA.

In attachment: the whereismycuda result, my Nebula configurations and the server panel.

nebula.JPG
nebula.JPG (78.05 KiB) Viewed 920 times


Thank you in advance for your replies.
Vernon
Vernon
User Level III
User Level III
 
Posts: 38
Joined: Wed Mar 31, 2010 3:27 pm
Location: Torino - Italy

Re: Again on CUDA performances

Postby enriquesilveti » Wed Jun 08, 2011 4:39 pm

Can you post here a screen shot of Nebula KERN page from the AlexB problematic emulation programs.
Enrique Silveti.
Acustica Audio customer and technical support.

MBP 11.5 (i7-4870 | 16 GB | 512 SDD)
SP4 (i5-6300 | 8 GB | 256 SDD)
RME UFX | PS Lyra2 | SD USBPre2
VM U15 | VM W10 | VM OSX 10.12
N4/NAT4 | SPX3 | RX5 | LN2C | Smaart8 | R5 | PT12 | PX10
User avatar
enriquesilveti
Expert
Expert
 
Posts: 2662
Joined: Sun Mar 28, 2010 9:00 pm
Location: Lodi | Madrid | Buenos Aires

Re: Again on CUDA performances

Postby giancarlo » Wed Jun 08, 2011 4:48 pm

ok, we'll repeat here again.

CUDA is for long tailed kernels.
Requisites:
1) kernels greater than 150 ms
2) huge performance gain if ALL kernels are tuned at the same kernel size
3) possibly kernels should not be dynamic


so you can use CUDA for
1) reverbs
2) static eq, where all kernels are longer than 150 milliseconds and possibly all kernels have the same length

for all other things you'll get poor performances, due to internal limitations. We are working on a different solution BUT we don't know about results yet.

BTW: you can use nebula/nebula server WITHOUT cuda... why are you using cuda for alexb presets? if you tune a proper server configuration with multiple lan connections, you could drive around 15 nebulas on each lan cable, and you could run many nebulas (someone reported 25 on i7) in localserver mode. I suggest the localserver/server solution because loading the same patch (alexb preset) you could save huge amounts of memory
User avatar
giancarlo
Founder
Founder
 
Posts: 9171
Joined: Mon Sep 21, 2009 10:40 pm
Location: Italy

Re: Again on CUDA performances

Postby Vernon » Wed Jun 08, 2011 7:12 pm

Thanks for the replies. In attachment the kernels used by Alexb emulation.

kernels.jpg
kernels.jpg (32.68 KiB) Viewed 887 times


giancarlo wrote:BTW: you can use nebula/nebula server WITHOUT cuda... why are you using cuda for alexb presets? if you tune a proper server configuration with multiple lan connections, you could drive around 15 nebulas on each lan cable, and you could run many nebulas (someone reported 25 on i7) in localserver mode. I suggest the localserver/server solution because loading the same patch (alexb preset) you could save huge amounts of memory


Well...I'm simply trying to get the most out of the resources I'm using. Of course I cannot use cuda for plugins other than Nebula.
And...of course I could drive 15/25 Nebula instances but I've to find space for other plugs that needs resources too.
When the Nebula3Server LAN version will be officially supported?

Thanks, Vernon
Vernon
User Level III
User Level III
 
Posts: 38
Joined: Wed Mar 31, 2010 3:27 pm
Location: Torino - Italy

Re: Again on CUDA performances

Postby giancarlo » Wed Jun 08, 2011 8:57 pm

ok. The good news: we are working very hard on cuda right now. You need to wait a bit. A configuration 50-50-50 is not good with cuda, unless you have a monster cuda card and a very slow CPU, which is not your case.

I'm using my new ipad only for reading nvidia/cuda documentation ;) porting the whole engine to cuda anyway is not a simple task, it requires a bit of time
User avatar
giancarlo
Founder
Founder
 
Posts: 9171
Joined: Mon Sep 21, 2009 10:40 pm
Location: Italy


Return to Working with Nebula

Who is online

Users browsing this forum: No registered users and 8 guests