Login

CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Tips & tricks, working results, technical support

CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby Prefect_X » Tue Aug 17, 2010 4:48 am

Following are the results from my recent experimentation with Nebula-3 Pro running through CUDA. I was unable to find any forum posts which discussed Nebula-3's CUDA performance, so I bought a GTS250card to unloading some of Reaper's plugin workload from the CPU, hopefully enabling more live/real-time editing scenarios.

My system was comprised of:

- 3.0Ghz AMD Phenom II X4 940
- Nvidia GTS250
- E-MU 1616m PCIe soundcard & breakout box
- Windows XP Pro x64 SP2
- no anti-virus or software firewall
- Reaper 64bit v3.66
- Nebula 3 Pro 64bit v1.3.413
- 64bit Nvidia Forceware driver v258.96
- 64bit Nvidia CUDA Toolkit v2.3 (latest version - v3.1 as of writing this - will not work for me)
- NO CUDA SDK -- the SDK is NOT required to run apps on CUDA cores, only for development!
- 64bit Acustica Audio CUDA Bridge v1.1.034


The test involved playing a 30 second long stereo wave file (recorded piano, if you must know) looped in Reaper with plugins applied (see below), then sequentially adding tracks until the occurance of audible artifacts or playback glitches (eg. clicking, popping, static, juttering, etc...). All resource numbers refer to the maximum PLAYABLE tracks for that test.

Reaper projects were set to 24bit/44.1Khz, & my ASIO buffer was set to 20ms. All Nebula-3 Pro plugins were loaded with default values of the "Dig Spring" patch, plus the master page was basically default values (including "DSPBUFFER = 8192") apart from switching between CPU and CUDA -- CPU used setting "OPT FREQD = 11 IPPS ch 2Mono" & CUDA used setting "OPT FREQD = 14 CUDA ch 2Mono". There are also non-Nebula-3 plugin references below to contrast against, namely the "Ambience" reverb by Smart Electronix (using default settings), & the "KORE-2" plugin by Native Instruments with default settings of the "Blind Marmelade" effect loaded.


Results:

(drum roll, please...)

==========

TEST 01 : 1 TRACK : NO PLUGINS

-- Task Manager reports 14% CPU load. Reaper reports 9.6% CPU load and 0% FX CPU load plus 34Mb of RAM used.

==========

TEST 02 : 1 TRACK : AMBIENCE REVERB PLUGIN

-- Task Manager reports 14% CPU load. Reaper reports 9.2% CPU load and 1.85% FX CPU load plus 45Mb of RAM used.

==========

TEST 03 : 1 TRACK : NEBULA-3 PRO ON CUDA ALONE

-- Task Manager reports 14% CPU load. Reaper reports 10.4% CPU load and 0.06% FX CPU load plus 249Mb of RAM used.

==========

TEST 04 : 1 TRACK : NEBULA-3 PRO ON CPU ALONE

-- Task Manager reports 16% CPU load. Reaper reports 11.5 % CPU load and 0.06% FX CPU load plus 247Mb of RAM used.

==========

TEST 05 : 1 TRACK : KORE-2 FX PLUGIN

-- Task Manager reports 20% CPU load. Reaper reports 10% CPU load and 7.5% FX CPU load plus 45Mb of RAM used.
-- Also noteworthy here is that Windows Task Manager shows a discrepancy of 111Mb, probably accounted for by the KORE-2 plugin currently loaded.

==========

TEST 06 : MAX TRACKS : NO PLUGINS

-- 190 tracks (after this I started noticing very minor glitches initially, but although I went to 250 tracks (and Reaper STILL didn't crash), by the 250 track mark playback was VERY scratchy and juttery.
-- Task Manager reports 43% CPU load. Reaper reports 37% CPU load and 0% FX CPU load plus 415Mb of RAM used.

==========

TEST 07 : MAX TRACKS : NEBULA-3 PRO ON CPU ALONE

-- 19 tracks (upon adding the 20th track, juttering and clicking immediately became quite bad)
-- Task Manager reports 97% CPU load. Reaper reports 94% CPU load and 2.89% FX CPU load plus 3886Mb of RAM used.

==========

TEST 08 : MAX TRACKS : NEBULA-3 PRO ON CUDA ALONE

-- 6 tracks (upon adding the 7th track, juttering and clicking immediately became quite bad, even worse than with the CPU test)
-- Task Manager reports 24% CPU load. Reaper reports 18% CPU load and 0.57% FX CPU load plus 1247Mb of RAM used.

==========

TEST 09 : MAX TRACKS : 6 X NEBULA-3 PRO ON CUDA + MAX ADDITIONAL TRACKS WITH CPU-POWERED NEBULA-3 PLUGINS

-- 3 extra tracks - total of 9 tracks playing (after setting up 6 tracks with CUDA powered plugins, I then added CPU powered plugins until I ended up with glitches, and surprisingly I could only add 3 more CPU powered tracks. The 4th CPU powered track with Nebula plugged in (10th track in total) once again cause major glitches with the playback.
-- Task Manager reports 37% CPU load. Reaper reports 29% CPU load and 0.98% FX CPU load plus 1860Mb of RAM use.

==========

TEST 10 : MAX TRACKS : AMBIENCE REVERB PLUGIN

-- 36 tracks (slight glitches with 37 tracks after several seconds of playing, but major glitches immediately with 38 tracks)
-- Task Manager reports 91% CPU load. Reaper reports 20% CPU load and 91.5% FX CPU load plus 127Mb of RAM use.

==========

TEST 11 : MAX TRACKS : 6 X TRACKS WITH NEBULA-3 PRO ON CUDA + MAX ADDITIONAL TRACKS WITH AMBIENCE REVERB PLUGIN

-- 6 extra tracks -- total of 12 tracks playing
-- Task Manager reports 39% CPU load. Reaper reports 22% CPU load and 42% FX CPU load plus 1281Mb of RAM used.

==========

TEST 12 : MAX TRACKS : 6 X TRACKS WITH NEBULA-3 PRO ON CPU + MAX ADDITIONAL TRACKS WITH AMBIENCE REVERB PLUGIN

-- 28 extra tracks -- total of 34 tracks playing
-- Task Manager reports 99% CPU load. Reaper reports 43% CPU load and 85% FX CPU load plus Mb of RAM used. Task Manager reports 1367Gb of RAM used.

==========

TEST 13 : MAX TRACKS : KORE-2 FX PLUGINS

-- 10 tracks
-- Task Manager reports 93% CPU load. Reaper reports 12.9% CPU load and 97.2% FX CPU load plus 67Mb of RAM used. Task Manager reports 994Mb of RAM used.

==========

TEST 14 : MAX TRACKS : 6 X TRACKS WITH NEBULA-3 PRO ON CUDA + MAX ADDITIONAL TRACKS WITH KORE-2 PLUGINS

-- 2 extra tracks -- total of 8 tracks playing
-- Task Manager reports 40% CPU load. Reaper reports 17.3% CPU load and 27% FX CPU load plus 1262Mb of RAM used. Task Manager reports 1.76Gb of RAM used.

==========

TEST 15 : MAX TRACKS : 6 X TRACKS WITH NEBULA-3 PRO ON CPU + MAX ADDITIONAL TRACKS WITH KORE-2 PLUGINS

-- 19 extra tracks -- total of 27 tracks playing
-- Task Manager reports 97% CPU load. Reaper reports 39% CPU load and 88% FX CPU load plus Mb of 1315RAM used. Task Manager reports 2.08Gb of RAM used.

==========

TEST 16 : MAX TRACKS : 2 X TRACKS WITH KORE-2 FX PLUGINS + MAX ADDITIONAL TRACKS WITH NEBULA-3 PRO ON CPU

-- 16 extra tracks -- total of 18 tracks playing
-- Task Manager reports 95% CPU load. Reaper reports 73% CPU load and 31% FX CPU load plus 3290Mb of RAM used. Task Manager reports 3.66Gb of RAM used.

==========

TEST 17 : MAX TRACKS : 2 X TRACKS WITH KORE-2 FX PLUGINS + MAX ADDITIONAL TRACKS WITH AMBIENCE REVERB PLUGIN

-- 29 extra tracks -- total of 31 tracks playing
-- Task Manager reports 95% CPU load. Reaper reports 21% CPU load and 99% FX CPU load plus 120Mb of RAM used. Task Manager reports 1.09Gb of RAM used.

==========

TEST 18 : MAX TRACKS : 2 X TRACKS WITH NEBULA-3 PRO ON CPU + MAX ADDITIONAL TRACKS WITH KORE-2 FX PLUGINS

-- 9 extra tracks -- total of 11 tracks playing
-- Task Manager reports 94% CPU load. Reaper reports 19% CPU load and 93% FX CPU load plus 478Mb of RAM used. Task Manager reports 1.31Gb of RAM used.

==========

TEST 19 : MAX TRACKS : 2 X TRACKS WITH KORE-2 FX PLUGINS + 6 X TRACKS WITH NEBULA-3 PRO ON CPU + MAX ADDITIONAL TRACKS WITH AMBIENCE REVERB PLUGIN

-- 16 extra tracks -- total of 18 tracks playing
-- Task Manager reports 99% CPU load. Reaper reports 77% CPU load and 28.5% FX CPU load plus 3292Mb of RAM used. Task Manager reports 3.66Gb of RAM used.

==========

Conclusion:

Using the GTS250 as my first foray into GPU-accelerated audio processing, I was ultimately able to run less simultaneous real-time instances of plugins while Nebula-3 Pro was (apparently) offloading it's workload to my CUDA enabled GTS250 graphics card.

So for all of you who were (like myself) excited about the potential for CUDA to be the perfect off-CPU number-crunching, effect-accelerating platform for real-time editing on your Digital Audio Workstation... Sorry. This disappointing Nebula-3/CUDA experience could simply be due to the trailblazing nature of Acustica Audio's project at the moment, and hence merely the tip of the iceberg for what the CUDA ideal can eventually bring to the table for audio processing, but for those of you looking for more multi-track, parralel-plugin-processing power RIGHT NOW... you probably need to keep saving for that Quad-G34 socket Tyan board matched with 12-Core Opterons to ease your suffering... like I am. Besides, until somebody (**cough** Reaper **cough**) has some CUDA-enhanced VST wrapper arrangement built into their DAW that will automatically redirect number crunching to your GPU, a multi processor board is still going to be the best way of improving ALL of your VST's performances, not just Nebula-3.

**swallows hard as he compares the latest 12-core Opteron prices on eBay with the balance of his credit card**

Considering the raw floating-point-number-crunching power of the graphics card, maybe Nebula-3's CUDA code just needs optimizing (or re-writing), or maybe there's a latency bottleneck to overcome, or maybe a GTX480 would have been more suited to the task at hand, but at least these tests (above) conclude that the GTS250 (or any other similar or lower spec'd GPU for that matter) is certainly not worth expending the dollars on -- at least not with the current state of Nebula-3's CUDA usage. Actually, considering the DECREASE in overall performance while utilizing Nebula's CUDA processing abilities, I may as well sell the GTS250 and go back to the fanless Radeon 4550 that was aiding that very expensive TRUE Copper cooler I shelled out for to increase "noise-free" performance aspect originally intended for my home studio recording computer.


Regards,


Dale Leighton
((Rig)): 3.0Ghz Phenom 940 Quad, 8Gb RAM, EMU 1616m PCIe & Roland VG-99.
((Soft Tools)): WinXP64, ReNoise, Reaper, Emulator-X3, EWQL Composers Collection, Addictive Drums, NI KORE-2 & Komplete-6.
User avatar
Prefect_X
User Level 0
User Level 0
 
Posts: 9
Joined: Sat Jun 19, 2010 1:35 pm
Location: Wellington NSW (Australia)

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby highvoltage » Tue Aug 17, 2010 1:06 pm

great test!
highvoltage
Member
Member
 
Posts: 235
Joined: Tue Mar 30, 2010 9:44 pm

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby ericus » Tue Aug 17, 2010 4:59 pm

Wow! Thank you for doing all those tests! :ugeek:
ericus
Member
Member
 
Posts: 243
Joined: Sun Mar 28, 2010 9:37 pm

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby giancarlo » Tue Aug 17, 2010 10:44 pm

you should test cuda 3.0 with nebula, performances are pretty good there. Could you?
User avatar
giancarlo
Founder
Founder
 
Posts: 9196
Joined: Mon Sep 21, 2009 10:40 pm
Location: Italy

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby Prefect_X » Wed Aug 18, 2010 12:04 am

I initially installed version 3.1 of the CUDA Toolkit, but it broke compatibility with Nebula on my computer, so after much mucking around, plus an email to for support to which you sent me back a cryptic response about needing a 64bit host (what was the go with that? I already said I had Reaper 64bit in my support email...???), I then tried going back to version 2.3 in a last desperate attempt to make my $120 (aus) worthwhile.

That is why v2.3 is the base for the tests...
((Rig)): 3.0Ghz Phenom 940 Quad, 8Gb RAM, EMU 1616m PCIe & Roland VG-99.
((Soft Tools)): WinXP64, ReNoise, Reaper, Emulator-X3, EWQL Composers Collection, Addictive Drums, NI KORE-2 & Komplete-6.
User avatar
Prefect_X
User Level 0
User Level 0
 
Posts: 9
Joined: Sat Jun 19, 2010 1:35 pm
Location: Wellington NSW (Australia)

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby giancarlo » Wed Aug 18, 2010 8:21 am

yes, the trouble is that you need cuda 3.0 and not cuda 3.1
User avatar
giancarlo
Founder
Founder
 
Posts: 9196
Joined: Mon Sep 21, 2009 10:40 pm
Location: Italy

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby Prefect_X » Wed Aug 18, 2010 1:19 pm

Just installed CUDA Toolkit 3.0 with high hopes of great performance, but alas, it failed to work... again.

:-(

From a programmer's perspective, it could be ANYTHING small and easy to fix, but the simple matter is; I can't compare if it doesn't work.

If you have a read of the second reply (by "yep") to the same post that I made on the Reaper forum:

http://forum.cockos.com/showthread.php?t=63505

...then I think a few of his points line up with what I was thinking. Good points I think. Not that I'm biased at all. **his eyes dart around the room in a guilty fashion**
((Rig)): 3.0Ghz Phenom 940 Quad, 8Gb RAM, EMU 1616m PCIe & Roland VG-99.
((Soft Tools)): WinXP64, ReNoise, Reaper, Emulator-X3, EWQL Composers Collection, Addictive Drums, NI KORE-2 & Komplete-6.
User avatar
Prefect_X
User Level 0
User Level 0
 
Posts: 9
Joined: Sat Jun 19, 2010 1:35 pm
Location: Wellington NSW (Australia)

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby enriquesilveti » Wed Aug 18, 2010 8:20 pm

While Samplitude is playing, I'm writing this:

WXP PRO SP2 x86 2600
N3P 1.3.430 x86 (no released yet)
Geforce Driver v 197.13 x86
CUDA Toolkit v 3.00 x86
AA CUDA bridge v 1.1036 x86
EP: X4 LEX91 Vocal Concert 441.n2v 360 MB (1K)
MAST PAGE: See SS
KERN PAGE: See SS
CUDA usage: 32% (total 8% each)
RAM: 2.27 GB
CPU: 18/22 %
Sequencer: Magix Samplitude 11 PRO
ASIO buffer: 1024
Sample Rate: 44.1 kHz

Mainboard: GigaByte GA-EP45-UD3R
CPU: Intel Core 2 Quad Q9400
CPU cooler: Asus Silent Square
RAM: Crosair GeIL Kit DIMM 8 GB DDR2-800 Quad (with pasive colling system)
PSU: NesteQ NA6201
OS HDD: Western Digital WD3200AAJB
Audio HDD: Western Digital WD5002ABYS (RAID)
CDR/DVDR: Plextor PX-716A
Video: Gigabyte Geforce 9600 GT
Audio: Tascam US-144
Case: Hightech HTCR 4U608
Fans: Papst 8412/NGLE (80x25x80) + Super Fan (80x25x80) x2
Airflow control: Zalman fan mate2
Attachments
CUDATEST.jpg
CUDATEST.jpg (613.05 KiB) Viewed 2057 times
Enrique Silveti.
Acustica Audio customer and technical support.

MBP 11.5 (i7-4870 | 16 GB | 512 SDD)
SP4 (i5-6300 | 8 GB | 256 SDD)
UFX | Lyra2 | USBPre2
VM U15 | VM Win10 | VM OSX 10.12
N4/NAT4 | SPX3 | RX5 | LN2C | Smaart8 | R5 | PT12 | PX10 | NIK5
User avatar
enriquesilveti
Expert
Expert
 
Posts: 2663
Joined: Sun Mar 28, 2010 9:00 pm
Location: Lodi | Madrid | Buenos Aires

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby giancarlo » Wed Aug 18, 2010 9:42 pm

ok, so the issue is different: cuda is not working in your system. cuda 2.3 was working pretty badly when kernel size was not big enough and mode 15 was not used and cpu was fast compared with gpu. One of advantages, you use ram in gpu, for example. I use cuda extensively now with vnxt plates, there are big performance boosts there. We'll try to fix your issues with cuda 3
User avatar
giancarlo
Founder
Founder
 
Posts: 9196
Joined: Mon Sep 21, 2009 10:40 pm
Location: Italy

Re: CUDA performance benchmarks with Nebula 3 Pro -- GPU vs CPU

Postby Prefect_X » Thu Aug 19, 2010 1:13 am

I think this line...

"ok, so the issue is different: cuda is not working in your system"

...is a bit of a broad overstatement. CUDA is DEFINITELY active on my system, as GPU-Z accesses it perfectly, however, if you meant that CUDA is not working for Nebula-3 on my system, then to be perfectly blunt, I was unaware that you had improved the CUDA performance through Toolkit 3.0 (something that isn't mentioned on the main website or CUDA related pages), and I was also only setting things up how the FAQs page and the Nebula wiki say to.

If you think there is something system specific about my machine/software setup that is causing the problem, just tell me what else you need to know so that I can furnish you with enough info to rectify the problem. One of the main reasons I bought Nebula-3 Pro was for the idea or freeing up those reverb cycles from my CPU. Not that I'd get rid of it now that I've used a while -- I like it -- but I did already have compressors, EQ's and reverbs that were doing the job, with varied levels of success.

:-)

-- Dale L
((Rig)): 3.0Ghz Phenom 940 Quad, 8Gb RAM, EMU 1616m PCIe & Roland VG-99.
((Soft Tools)): WinXP64, ReNoise, Reaper, Emulator-X3, EWQL Composers Collection, Addictive Drums, NI KORE-2 & Komplete-6.
User avatar
Prefect_X
User Level 0
User Level 0
 
Posts: 9
Joined: Sat Jun 19, 2010 1:35 pm
Location: Wellington NSW (Australia)

Next

Return to Working with Nebula

Who is online

Users browsing this forum: No registered users and 5 guests