Login

Cuda

Tips & tricks, working results, technical support

Re: Cuda

Postby giancarlo » Tue May 20, 2014 7:11 am

reverbs will benefit. It could happen you get performance boost for some other preset too but dont build a cuda server for things different from reverbs
User avatar
giancarlo
Founder
Founder
 
Posts: 9169
Joined: Mon Sep 21, 2009 10:40 pm
Location: Italy

Re: Cuda

Postby XanomalieX » Tue May 20, 2014 9:55 am

Debutante wrote:Hi everyone...

I'm going to get a new build and perhaps upgrade my Nebula specifically to exploit this Cuda advantage, and I have some very general questions..

- how difficult is it to get going

- has anyone got any real world figures (so many instances of this preset type thing) for just how well Cuda works.

I've read every post in this topic and it seems the message is a little mixed, especially on what presets can benefit most from Cuda.

I'm a way better than average computer user, and not scared of Nebula, so I'm not afraid to learn a few more things, but it almost seems like an enthusiasts endeavour as opposed to a musician's one. Just how worth it is it?


It's really easy to set up and runs stable even with the newest version 6.
But don't expect something like dsp cards, with nearly no CPU usage at all!!

On my system: i7 2600 16GB RAM
I opened 16 nebula EAR Reverbs with 4096 buffer size
and I got:
90% CPU usage (just CPU)
60%-65% CPU usage with Cuda

it's nice but I'll never use 16 Reverbs, maybe 2 Rooms 1 Plate and 1 Reverb = 4 in one song
User avatar
XanomalieX
User Level IV
User Level IV
 
Posts: 48
Joined: Thu Dec 16, 2010 12:20 am
Location: Germany

Re: Cuda

Postby markrandallpixley » Sat May 31, 2014 4:21 am

So just thought I'd chime in here, I finally figured out how to get the free version to run Cuda I am on an Athlon quad-core with only 8 gigs of Ram...but I am running a GTX 470 video card that has 480 cores...in Sonar X3 (everything is 64 bit/OS/Sonar/Nebula)I was able to run 25 instances of the free reverb program and in my task manager it only it 50-60% of actual cpu (which is to be expected )...no drop-outs, no glitches, no headaches (well other than getting the setup to work in Cuda it kept giving me a flashing "!" until I moved around some dll's)...

At any rate this is incredible and I think the future I will definitely be purchasing a full-blown version and creating some of my own files (we have a small boutique audio gear company where we make vintage 1176's and a REDD 47 pre-amp as well as A*I/N**e/Sony/Urei/Pultec/Telefunkin rebuilds etc...I will definitely be busting out the parts and finishing the LA2A as well...so people can have actual signal chains of these great vintage products)...

I will keep expereimenting but this is really great!

Summary:

25 instances of Nebula
25 Instances of reverb on Cuda
CPU at 50-62% (real world)
30% in Sonar

Athlon Quad Core 2.8
8 gigs Ram
Geforce GTX 470 dual moniters
Cuda 6
Nebula free 64


I could probably do more but this is enough for me to see what it is capable of...

By the way I own and run a UAD-1 pci-e on this same machine and there is no way it could do this...
markrandallpixley
User Level I
User Level I
 
Posts: 13
Joined: Tue May 27, 2014 7:07 am

Re: Cuda

Postby jorismak » Sat May 31, 2014 10:48 am

I'd be interested to know what kind of programs you uses because I can't get past 2 or 3 instances.

I know the old 4xx serie (specially the 470 and 480) are way more efficient in cuda and gpgpu stuff than the newer series.

But I thought the newer series fixed that by sheer raw power and number of gpu blocks :).
jorismak
Member
Member
 
Posts: 344
Joined: Fri Nov 16, 2012 4:49 am

Re: Cuda

Postby markrandallpixley » Sat May 31, 2014 7:32 pm

I'm running Sonar X3 Producer, I've been using Cakewalk for nearly 15 years, so I upgrade as a matter of habit with them...

I run three instances of nebula/verb in one track and then clone it 8 times (also had one more track with a single instance)...

To be fair this machine is not optimized for audio much, I was also online at the time...however I am running two separate SSD drives one for my main OS and one specifically for audio...

If I was going to make a suggestion besides more RAM it would definitely be a new SSD drive...and a good video card...this card has 480 cuda cores...the newer ones are like 2800....something ridiculous...

I am not a big plug-in user since I have some really great vintage outboard gear, but the plug-ins here are the first ones to really catch my attention...I did manage to get the sound to break up after about 30 instances...

I'm running Lynx Audio converters (in my opinion best bang for the buck, we've used Apogee and Prism and Lynx holds its own and better on pocket book and service)

I was also running two instance of Kontakt (no instruments loaded though)Stylus as well and an instance of a vst3 meter plug-in...so there was a lot of memory demands on Sonar, but it handles vst's quit well...
markrandallpixley
User Level I
User Level I
 
Posts: 13
Joined: Tue May 27, 2014 7:07 am

Re: Cuda

Postby jorismak » Sat May 31, 2014 7:49 pm

What does 'cloning' the track mean? Isn't that just 3 instances and the output duplicated 8 times ( and thus are still 3 instances ) ??

Yes, newer cards have crazy shader counts. But Kepler needs more cores for the same FP-calculations. When the 680 came out it was winning in games, but gpgpu stuff was WAY faster on the older 580 and 570.

They've slowly worked on that so I don't know how it is now. Just don't compare shader count for shader coun, gpu architecture differs a lot :) .

But still checking how you get to that many reverb instances . I wanna know if my setup is under performing :).
jorismak
Member
Member
 
Posts: 344
Joined: Fri Nov 16, 2012 4:49 am

Re: Cuda

Postby markrandallpixley » Sat May 31, 2014 8:48 pm

No in Sonar "cloning" places all the audio and effects/sends/settings into a completely new track so when I "clone" that many instances I have to wait about 30 seconds for Sonar to load and insert 8 new audio tracks (using the existing wav file) and also load 3 new instances of the Vst (Nebula with a a free reverb hall I think it was) into each track...all getting sent to whatever insert bus and conditions I had set in the original track...it's a rather handy feature for me to test vst's with since I can clone a track and "print" (render/bounce) a track (Sonar will do this in realtime with a 64 bit engine also offline if I choose, which is really handy with stuff like Kontakt where I can actually bounce to an audio track a vst intrument and now I have an actual wav version of the cello/synth/whatever I was running)...

So these are all new tracks simply using the same settings throughout...

I was also able to get the midi module in Sonar to control the parameters on Nebula eq's...which is great since the sliders are non-tactile and extreme at times...

Sonar does run better on Intel chips, but I have been building my own systems for so long it's hard fro me to not use AMD because of cost...

My buffers were set at 1024 in both Nebula and Sonar and Asio and I also tweaked the "rate conv." in Nebula to 9000 ms...it is running Opt. Freqd: Cuda 15 ac2mono

My NAT setting and Sonar both agree that a "roundtrip" buffer takes about 72 ms...


So here's what it looks like:
Win 7 64
Sonar X3 Producer 64
Amd 2.8 Quad with 8 gigs DR3
2 SSD harddrives
Nebula free 64


Lynx Asio 1024/
Sonar 1024/
Nebula DSP buffer 1024/
Nebula Opt FreqD 15Cuda ac 2 MONO
Nebula rate conv. 9000 ms (highest possible)
Gui rate 30 ms
Kernals 10 (9th harm)


An i5 is probably comparable to this computer...although this particular AMD has 6 megs L3 buffer cache and makes all ram available for all cores...(a bios setting for me)

(Sonar is set at 23 ms playback so with this many instances it does "pause" about 3 seconds before start or stop but plays fine)

I will say this about the Lynx cards, the soundcard is an important part of this chain, I have used MAudio/Motu/Creative/Emu/Prism/Apogee and Echo and without question the way the company codes the card for Asio is huge...Lynx stays on top releasing new drivers every quarter (at least) also running the latest nVidia Cuda drivers and card drivers is huge if you are wanting to do this...

(One other thought added: Sonar shows a 30% CPU usage, but that is not accurate, Sonar shows what is available to it, not the actual, even when I have set it's affinity to use all 4 cores, it only ever uses 3...same is true with Nebula, only 3 cores were ever used...I know this by pulling up Windows Task manager during the playback and loading the system monitor to actually see...it was running pretty heavy at 62% at times but always had one core idle...not sure why that is, but it has always operated that way, maybe a an AMD thing I don't know)
markrandallpixley
User Level I
User Level I
 
Posts: 13
Joined: Tue May 27, 2014 7:07 am

Re: Cuda

Postby jorismak » Sun Jun 01, 2014 1:27 am

the cloning and bouncing you describe sounds like basic functionality in any DAW :).

Got curious, did a bit of testing.
44khz, 1024 ASIO buffers.
Had a single trick with some audio on it, first a limiter (bx_limit) to raise the level without making it peak, then 3 instances of my 'nebula3 cuda' (with the same 'mode 15' as you have).

Made copies of that trick will it started to become issues.

At 7 tracks (7x3 instances) it was just on the brink of working / having pops. Cpu load in Windows task manager was at 95% at that time.
Taking 6 tracks (so 6x3 instances) dropped CPU load to +/- 85% and caused everything to playback perfectly.

This is on a Core i7-860 @ 3.37ghz, which should be _leaps_ faster than your AMD, but on a GTX 760. Which proves that it's doing something on the card since my CPU should be better, but my gpu is a bit slower I think.

In games my Gtx760 would smoke, but for plain brute Cuda / GPGPU performance your older Fermi Gtx470 is still faster, and I think this kinda proves that :).

Still, it depends heavily on the program loaded. What kind of harmonic kernels and how long the reverb tail is. I used the Acustica program 'old plate 3', which is a 2-kernel EMT140 sampling of +/- 3.5seconds if I read the kernel-tab correctly.

I got my Nebula DSP buffer on 4096 for the Reverb/Cuda instances, giving me a total latency of 28672 samples on a single track (just over half a second). So pressing 'play' only gives me about a 0.5 second delay before Reaper starts playing.

I only don't know if I'm using Cuda5 or Cuda6 at the moment to be honest, I've been testing both. I am using Nebula3 cuda bridge 1.1.040, which I think is still the latest.

I did the same test with the CPU-only instance. There it was unworkable with 6 tracks (6x3 = 18 instances) but worked fine with +/- 80% to 86% CPU usage with 5x3 = 15 instances.

So in my case CUDA does help, but only a very little bit :).
But this is with a good fast CPU to be honest. If you have less cpu power (like you) you'll see more benefit. And it is with a simple 'mid range' GPU card.
Getting a Titan or a real Quadro-workstation card based on a 680 / Titan / 780 will boost the numbers I guess.
Or get an older 580 for cheap since they are only good for number-crunching like this :P.

I have the feeling that people with real entry level cards (610 / 620 / 730) will see little benefit. If you have a real slow cpu you might have too little CPU power to 'drive' your video card, and having a cheap video card will not help that much I guess.

For people with quad-core AMD cpu's or Intel i3 / i5 it might be just that little boost to get a few more reverb aux's in your mixing template, in which case it'll be great. But for people with recent Core i7's (or older i7's clocked higher :P) it doesn't help _that_ much... but it does help :).
jorismak
Member
Member
 
Posts: 344
Joined: Fri Nov 16, 2012 4:49 am

Re: Cuda

Postby markrandallpixley » Sun Jun 01, 2014 6:03 am

Really good points, I bought the 760 card myself from Amazon thinking it would boost performance (I think my windows index thingy was at 3.6 bottle-necked by the gpu) so I was watching craislist when this card popped up for $50.00 I thought what the heck, immediately my performance index jumped to 7.2 on Windows and even though it was an older card I saw the specs were really good so I called the guy and bought it...it has dual dvi outs and I run 2 monitors so that was cool...I don't do games but stood and chatted with the guy about what he does and he's always watching nvidia to catch the latest...but they are sort of hit and miss in what they put out...I paid $49.00 for the newer GTX 760 but it has less cores than this older card...

By the way make sure "splith" is emabled as well...I will never run that many instances of a reverb (we have an actual plate reverb in the shop we are building) I have some very good outboard verbs as well...but we are looking to sample some stuff now...so we will probably upgrade a machine and vid card as well...
markrandallpixley
User Level I
User Level I
 
Posts: 13
Joined: Tue May 27, 2014 7:07 am

Re: Cuda

Postby giancarlo » Sun Jun 01, 2014 10:55 am

cuda needs long kernels and fft ones. A limiter is the worst starting point possible. Make sure all kernels are the same size possibly and longer than 110 ms at 44100 or longer than 250 at 96000. Reverbs and distorted reverbs (the common things who destroy an i7 processor) are good starting point
User avatar
giancarlo
Founder
Founder
 
Posts: 9169
Joined: Mon Sep 21, 2009 10:40 pm
Location: Italy

PreviousNext

Return to Working with Nebula

Who is online

Users browsing this forum: No registered users and 6 guests