Targeting Nividas CUDA as processing unit of the future?
Targeting Nividas CUDA as processing unit of the future?
Sonic cores future Begins new with a new processing unit CUDA?
I just tested out using convolution based effects on music production with CUDA and an 8800 GPU. Man that thing is a HUGE number cruncher! A reverb that usually takes around 20% of my dualcore CPU eats less than 1% of the GPU.. I could load up dozens(maybe hundreds?) of them at the same time. That is definetly the the future path for audio processing. Not sure tho how other algorithms will sit on such GPUs, but even if their theoretical maximum speed can't be reached they will still be a good addition to the CPU. Running only convolution/fft stuff on them is also a huge benefit as it's ofcourse a huge reduction to the main CPU load.
Now that those old 8800(or 8600 etc) GPU's are dirt cheap i can see a huge market for VST's utilizing their power
BUT: Zero latency mode is not supported by the GPU Edition due to a combination of factors including the mechanisms involved with transferring data to and from the GPU being much more efficient with larger blocks, a current requirement to run CUDA VST plug-ins in a separate thread (making larger block processing more efficient) and the lack of coherency in buffering schemes used by various different VST hosts complicating the above issues. It may become more practical to implement this feature in future using newer GPU architectures.
I think the odds of me porting to CUDA are about zero.
The reason for this: GPUs are geared towards massively parallel processing. This is the sort of thing that you see in images, which are essentially NxM arrays of data. Image processing tends to be feedforward, and not reliant on previous outputs of the filters.
Most of the audio processes I work on are based on delays, feedback processes and filters, that sort of thing. This doesn't work well in massively parallel systems. I process audio in parallel whenever I can get away with it, for efficiency, but the fact is that delay lines of different lengths are inherently difficult to parallelize. Feedback doesn't work well with massively parallel systems, unless you have feedback with pretty big blocks, and that tends to lose a lot of the cool aspects of feedback.
Convolution processing (including Nebula) works well with massively parallel processing. Beyond that, I could see GPUs used for other things that work well with lots of parallelism, like modal synthesis. For general audio applications, I don't think that GPUs will prove that useful. I wouldn't mind being proven wrong, as there is a LOT of power in the GPU.
I just tested out using convolution based effects on music production with CUDA and an 8800 GPU. Man that thing is a HUGE number cruncher! A reverb that usually takes around 20% of my dualcore CPU eats less than 1% of the GPU.. I could load up dozens(maybe hundreds?) of them at the same time. That is definetly the the future path for audio processing. Not sure tho how other algorithms will sit on such GPUs, but even if their theoretical maximum speed can't be reached they will still be a good addition to the CPU. Running only convolution/fft stuff on them is also a huge benefit as it's ofcourse a huge reduction to the main CPU load.
Now that those old 8800(or 8600 etc) GPU's are dirt cheap i can see a huge market for VST's utilizing their power
BUT: Zero latency mode is not supported by the GPU Edition due to a combination of factors including the mechanisms involved with transferring data to and from the GPU being much more efficient with larger blocks, a current requirement to run CUDA VST plug-ins in a separate thread (making larger block processing more efficient) and the lack of coherency in buffering schemes used by various different VST hosts complicating the above issues. It may become more practical to implement this feature in future using newer GPU architectures.
I think the odds of me porting to CUDA are about zero.
The reason for this: GPUs are geared towards massively parallel processing. This is the sort of thing that you see in images, which are essentially NxM arrays of data. Image processing tends to be feedforward, and not reliant on previous outputs of the filters.
Most of the audio processes I work on are based on delays, feedback processes and filters, that sort of thing. This doesn't work well in massively parallel systems. I process audio in parallel whenever I can get away with it, for efficiency, but the fact is that delay lines of different lengths are inherently difficult to parallelize. Feedback doesn't work well with massively parallel systems, unless you have feedback with pretty big blocks, and that tends to lose a lot of the cool aspects of feedback.
Convolution processing (including Nebula) works well with massively parallel processing. Beyond that, I could see GPUs used for other things that work well with lots of parallelism, like modal synthesis. For general audio applications, I don't think that GPUs will prove that useful. I wouldn't mind being proven wrong, as there is a LOT of power in the GPU.
out and about for music production. Are you still configguring your Studio music first!
Re: Targeting Nividas CUDA as processing unit of the future?
Anabella you took these messages from a variety of sources, yeah?
Like this from posts to the synthmaker forum in 2008:
http://synthmaker.co.uk/forum/viewtopic ... 1&start=15
From the LiquidSonics site on their Reverberate reverb VST:
http://www.liquidsonics.com/software_reverberate_le.htm
And quotes from Sean Costelloe, developer of Valhalla VSTs:
http://www.gearslutz.com/board/product- ... ws-12.html
It's best to quote your sources in order to get the full picture...
As for Sonic Core's intention to port their atoms to a different hardware platform, for me it's more important that they devote resources to getting Scope 6 complete, improving the DSP placement and optimisations discovered by advanced SDK users here on planetz as part of default techniques employed by the platform, squashing the minor irritant bugs, and then getting a Mac-compatible version of Scope going.
There's nothing wrong with running Scope on SHARCs. The margins on hardware is more likely to gain S|C more than the sale of the software devices. And requiring S|C hardware is an effective solution to copy protection too.
Like this from posts to the synthmaker forum in 2008:
http://synthmaker.co.uk/forum/viewtopic ... 1&start=15
From the LiquidSonics site on their Reverberate reverb VST:
http://www.liquidsonics.com/software_reverberate_le.htm
And quotes from Sean Costelloe, developer of Valhalla VSTs:
http://www.gearslutz.com/board/product- ... ws-12.html
It's best to quote your sources in order to get the full picture...
As for Sonic Core's intention to port their atoms to a different hardware platform, for me it's more important that they devote resources to getting Scope 6 complete, improving the DSP placement and optimisations discovered by advanced SDK users here on planetz as part of default techniques employed by the platform, squashing the minor irritant bugs, and then getting a Mac-compatible version of Scope going.
There's nothing wrong with running Scope on SHARCs. The margins on hardware is more likely to gain S|C more than the sale of the software devices. And requiring S|C hardware is an effective solution to copy protection too.
Not because it is easy, but because it is hard...
Re: Targeting Nividas CUDA as processing unit of the future?
Well, if SC were going to port he atoms to anything, it would be the chip they are already running on in XITE (21369) wouldn't it ?
That would yield far greater performance / benefit than any other optimisation or DSP placement...after all, if you can run more on a single chip then placement becomes less off an issue anyway.
That would yield far greater performance / benefit than any other optimisation or DSP placement...after all, if you can run more on a single chip then placement becomes less off an issue anyway.
Last edited by dante on Sun Jan 26, 2014 3:04 am, edited 1 time in total.
Re: Targeting Nividas CUDA as processing unit of the future?
good recherche! My english isnt so good. Ok keep the scope for some exclusive users. But if cuda will not come, maybe an other comes. For me, the scope is perfect, but very expensive.
out and about for music production. Are you still configguring your Studio music first!
Re: Targeting Nividas CUDA as processing unit of the future?
That makes even less sense. Porting it to another totally different chip (cuda) would mean paying for years more R&D than a full 21369 port, making Scope way more expensive than what it is now.
I'm not sure how you reach the conclusion that Scope is expensive. I have rack mounted mixer, multiband compressor and reverb the retail price that together cost me more than what the XITE1-D cost, yet the XITE already performs the same function of those 3 units and at least 3 more besides simultaneously.
I'm not sure how you reach the conclusion that Scope is expensive. I have rack mounted mixer, multiband compressor and reverb the retail price that together cost me more than what the XITE1-D cost, yet the XITE already performs the same function of those 3 units and at least 3 more besides simultaneously.
Re: Targeting Nividas CUDA as processing unit of the future?
i have a two channel mic pre that cost as much as an XITE-1.
i have a mixing board that was some $28,000 in 1987 or so...
i have a mixing board that was some $28,000 in 1987 or so...
Re: Targeting Nividas CUDA as processing unit of the future?
I first remembered these Berkely? grad students being purchased by NVidia when they first came out in the latter end of the '90s.
They did a Reverb BETA testing thing online and had Reverb FX stomping all over the Lexicon PCI Card, NuVerb and Steinbergs Metallic Hall (as we called it).
NVidia has basically kept this sweet stuff from people for a long time.
Back then NVidia, Matrox and ATI AGP 4x stuff had more power than the Celeron 200As and ScopesDSPs.
They did a Reverb BETA testing thing online and had Reverb FX stomping all over the Lexicon PCI Card, NuVerb and Steinbergs Metallic Hall (as we called it).
NVidia has basically kept this sweet stuff from people for a long time.
Back then NVidia, Matrox and ATI AGP 4x stuff had more power than the Celeron 200As and ScopesDSPs.
Re: Targeting Nividas CUDA as processing unit of the future?
I often thought that the idea of applying GPU's, as massively-parallel processors with a substantial memory allocation, to the processing of audio signals, is theoretically a great one. I do wonder why they're not used, why there isn't a dedicated Audio SDK produced to leverage their capabilities for audio processing...
DSP is DSP, no matter what the Signal represents (audio or video), I would have thought.
I often wondered too why no one tried doing a D/A converter that took DVI or HDMI as digital input - instead of USB...
Even if the GPU were loaded with just a Convolution Processor - we've all read about or experienced what Nebula can do - we could have something like a quality rackmount effects unit inside our PCs.
But I'm no DSP programmer! I really have nothing of value to add to a debate on technical deficiencies or difficulties - feedforward/feedback algorithms, delay lines, how they would interact with the capabilities of a graphics card, and who knows what else... So I'll stay quiet!
DSP is DSP, no matter what the Signal represents (audio or video), I would have thought.
I often wondered too why no one tried doing a D/A converter that took DVI or HDMI as digital input - instead of USB...
Even if the GPU were loaded with just a Convolution Processor - we've all read about or experienced what Nebula can do - we could have something like a quality rackmount effects unit inside our PCs.
But I'm no DSP programmer! I really have nothing of value to add to a debate on technical deficiencies or difficulties - feedforward/feedback algorithms, delay lines, how they would interact with the capabilities of a graphics card, and who knows what else... So I'll stay quiet!
Not because it is easy, but because it is hard...
-
- Posts: 1638
- Joined: Mon Nov 15, 2010 12:57 pm
Re: Targeting Nividas CUDA as processing unit of the future?
Audio DSPs have a certain set of instructions used for calculating audio. It's what makes them so efficient (ie-why a 33mhz DSP can have several voices on it for example).
The GPUs don't have these instructions, so they would have to use lots of instructions to recreate the same algorithm. What these massively parallel processors don't have is a large instruction space. For convolution, the algorithm is very similar to what is done with video, and the massive parallelism addresses the huge amount of data a convolution reverb calculates.
If all of the DSP instructions were added to a GPU, each core would get much bigger, and so the instead of having 256 processors on a chip, you might get a handful, and the frequency would be lower also.
Further, the biggest task to deal with in the XITE, is shuffling the data around between the DSPs, which can force SAT connection errors. The GPU is worse for those connections, it is optimized for data to general flow in blocks and in a specific direction. It wouldn't work very well for sample-accurate transfers and you'd end up with higher latency like Native deals with.
The Creamware guys did a very good job addressing the issues of how to process so much audio in real time. Think about how UA has an almost identical hardware design. It's because it works well.
The GPUs don't have these instructions, so they would have to use lots of instructions to recreate the same algorithm. What these massively parallel processors don't have is a large instruction space. For convolution, the algorithm is very similar to what is done with video, and the massive parallelism addresses the huge amount of data a convolution reverb calculates.
If all of the DSP instructions were added to a GPU, each core would get much bigger, and so the instead of having 256 processors on a chip, you might get a handful, and the frequency would be lower also.
Further, the biggest task to deal with in the XITE, is shuffling the data around between the DSPs, which can force SAT connection errors. The GPU is worse for those connections, it is optimized for data to general flow in blocks and in a specific direction. It wouldn't work very well for sample-accurate transfers and you'd end up with higher latency like Native deals with.
The Creamware guys did a very good job addressing the issues of how to process so much audio in real time. Think about how UA has an almost identical hardware design. It's because it works well.
Re: Targeting Nividas CUDA as processing unit of the future?
Yep, I appreciate that the raw DSP of a SHARC chip is very different in its intent to a graphics processor. There isn't so much Mathematics in Graphics as there is in the subtleties of analog-emulating oscillators, for example.
But straight convolution, which can do lots to audio, and works on 'batches' of data (I only know how Blur, Sharpen, Despeckle type convolution processors work!), would be an efficient use of the capabilities of 2D audio spaces?
I also appreciate that the GPU works best when given large blocks of data on which to do their simple operations, which compromises their ability to process streams efficiently in near-real-time, which therefore relegates their usability in the audio context to the second division...
But you would think that you could do one heck of an additive synth in such a domain too... any number of partials, all calculated in parallel..... that'd make for one crazy vocoKder too....
Real-time wavetables, or granular synthesis, are also candidates, you'd think.
Imagine installing a second graphics card, taking a DVI out to a rackmount unit with D/A into Analog filters, with high-resolution cv-style control over analog envelopes and lfo components....
Yep I am probably talking thru my hat.......
Did I say I'd stay quiet?!
I definitely am not offering the idea that Scope could be ported to a CUDA-like system. It simply is not a good technological fit. GPU's do one thing well, Scope does loads of things well.
But straight convolution, which can do lots to audio, and works on 'batches' of data (I only know how Blur, Sharpen, Despeckle type convolution processors work!), would be an efficient use of the capabilities of 2D audio spaces?
I also appreciate that the GPU works best when given large blocks of data on which to do their simple operations, which compromises their ability to process streams efficiently in near-real-time, which therefore relegates their usability in the audio context to the second division...
But you would think that you could do one heck of an additive synth in such a domain too... any number of partials, all calculated in parallel..... that'd make for one crazy vocoKder too....
Real-time wavetables, or granular synthesis, are also candidates, you'd think.
Imagine installing a second graphics card, taking a DVI out to a rackmount unit with D/A into Analog filters, with high-resolution cv-style control over analog envelopes and lfo components....
Yep I am probably talking thru my hat.......
Did I say I'd stay quiet?!
I definitely am not offering the idea that Scope could be ported to a CUDA-like system. It simply is not a good technological fit. GPU's do one thing well, Scope does loads of things well.
Not because it is easy, but because it is hard...
Re: Targeting Nividas CUDA as processing unit of the future?
So we win twice by having SHARCs BC Modular, and SHARC DSPs.......
ANKYU..
ANKYU..
Re: Targeting Nividas CUDA as processing unit of the future?
You can go both directions, I have a 499€ PC with a behringer soundcard 49€ and cubase 7 and it does make musicproduction possible, better than ever before with any Hardware.
But ok, I know best, what vast Quality the Pulsar platform offers, and why People get laufing when they hear VST, so ok, it is a very very good Price.
Please dont get angry to me, because I said, SC isnt cheap, but it is very much for a lot of Money. In relation Price and what you get, ok it is a good relationship.
But ok, I know best, what vast Quality the Pulsar platform offers, and why People get laufing when they hear VST, so ok, it is a very very good Price.
Please dont get angry to me, because I said, SC isnt cheap, but it is very much for a lot of Money. In relation Price and what you get, ok it is a good relationship.
out and about for music production. Are you still configguring your Studio music first!
- Bud Weiser
- Posts: 2709
- Joined: Tue Sep 14, 2010 5:29 am
- Location: nowhere land
Re: Targeting Nividas CUDA as processing unit of the future?
Thinking professional,- you don´t get any hi-end PC or laptop for EUR 499.- (incl. VAT !!!) and a Behringer soundcard for EUR 49.- is also not the ice on the cake.anabella wrote:You can go both directions, I have a 499€ PC with a behringer soundcard 49€ and cubase 7 and it does make musicproduction possible, better than ever before with any Hardware.
It´s USB,- right ?
And,- in addition for what it does,- how long will "it" last ?
I say "it" because it´s not only one piece of gear but a bunch of several components all w/ a different MTBF.
No, it´s not a very good price but it is a very good deal for the manufacturers.anabella wrote: But ok, I know best, what vast Quality the Pulsar platform offers, and why People get laufing when they hear VST, so ok, it is a very very good Price.
See Cubase (I was a Cubase user) ...
For every update you pay and that rules for every single commercial native plugin.
At some point, upgrading is impossible,- I couldn´t upgrade Wavelab anymore because I used v4.01b for too long and didn´t pay all year,- I was pi##ed !
When I decided to upgrade from Cubase SX3, it was also too late because the upgrade alone cost EUR 499.- then.
That was the time to buy Presonus Studio One Pro for half the price and have more features and better workflow.
Then I discovered Reaper for 60 bucks,- well ...
Native plugins and applications will be discontinued one day because the code is bloated and they have to rewrite all.
You lose.
Sometimes plugins are discontinued just only because they don´t want to support ´em anymore and want you to buy new stuff.
The gross profit margin for native software is 60%+ and the most expensive production cost is the box, box design, booklet documentation and put the CD or DVD into cellophane.
If it is a download product, less costs.
A former europaen sales manager of a leading music software company is my friend, so I now some details.
With every upgrade of native software,- the demand for a faster computer increases.
You have to buy (or build) a new machine much more often when using native VST/VSTi (or AU) only.
When the computer hardware specs change,- connections disappear p.ex.,- you need a new soundcard and today it seems only USB survives.
You always pay what you get for.anabella wrote: Please dont get angry to me, because I said, SC isnt cheap, but it is very much for a lot of Money. In relation Price and what you get, ok it is a good relationship.
The old Creamware cards are very reliable equipment and you find enough old computers today to use the cards.
Ebay is full of used HP servers ... and you´re able to upgrade SCOPE even the cards come from the 90s,- isn´t that great ?
As a hardware device, the XITE-1 or 1D are extremely well built pieces of hardware gear.
Any UAD card or box using fewer new SHARC chips isn´t cheaper in relation to it´s DSP power.
In fact, when you buy a UAD Apollo incl. all the plugins available,- it is significantly more expensive than a XITE-1,- and the Thunderbold connection is optional and not included.
When you buy XITE-1, you get latest SCOPE and all devices except 4 synths (ProTone, Prodyssey, MODIV and Sixstring),- but you have a Modular III synth in SCOPE, small & big synths, samplers,- as well as all FX S|C developed.
3rd party devices are out of discussion here.
I´m pretty sure you can do almost any music w/ stock SCOPE alone and a decent PC which runs your DAW application and a sampleplayer device delivering the bread & butter sounds like pianos, e-pianos, other acoustic and electromagnetic instrument emulations etc.,- say NI Kontakt or use Phead Reason as you DAW app and refills.
I also don´t see much or any virtual instruments devices for UAD Apollo.
You should be more carefull what you´re comparing when typing posts.
It´s correct native VST/AU stuff is fine for most users and (paying ?) customers of music because most mixes end in a 16Bit/44,1K file or mp3 anyway and will be listened to from cellphones and w/ earbuds,- but there´s also a reason why there is lots of hardware in recording and mastering studios,- outboard and consoles as there is a reason why DSP hardware is existing.
VST/AU only is for the home users, bedroom producers and amateurs.
I don´t know any pro studio in my hometown using native software only.
I have to say they use Apple Logic and Mac Pro computers more often that PCs or have both,- PC and Mac,- and as Mac users they will decide for UAD Apollo w/ Thunderbold when looking for DSP, because there´s no OS X compatibility of SCOPE/XITE up to now.
It´s speculation,- but it also could change w/ SCOPE 6 being OS X compatible one day.
just my opinion ...
Bud
S|C Scope/XITE-1 & S|C A16U, Scope PCI & CW A16U
Re: Targeting Nividas CUDA as processing unit of the future?
out of all dsp devices there are only 2 with free plugs
chameleon
and scope
all other dsp systems you had to pay for it
chameleon is no more tc dsp card support or development is no more
uad new plugs cant run on old cards
the scope system is the only one that can play on old tech and new
so many free dsp modular modules and new ones being built all the time
at least scope systems had newer drivers to work with newer machines
oasys dsp card is stuck on win98 its a nice card and you can do some great things on it but only on a old mac or pc
not many synths on other dsp platforms
scope is not expensive at all i payed £10000 +£4000 +£3000 for an EII a macII and a soundtools and sd2 in 1985
now that was expensive and for 400k banks
resale value for that now is not much above a £1000
the cards are cheap in comparison and do much more than i could ever dream of
my first card i brought in 2000 and its still going strong 14 years later still supported with new soft and new plugs
my protools system which i paid £10000 for and a few 1000,s for interfaces became defunkt when they went from pci-x to pci-e and the upgrade path was pitance but if you did not upgrade then your system was none supported then they ditched it for native so i was glad i never upgraded
but with mac systems as soon as they changed the comp spec your old gear if you was upgrading was obsolete
but not the pci cards from scope i have upgraded several times and still my cards perform well
chameleon
and scope
all other dsp systems you had to pay for it
chameleon is no more tc dsp card support or development is no more
uad new plugs cant run on old cards
the scope system is the only one that can play on old tech and new
so many free dsp modular modules and new ones being built all the time
at least scope systems had newer drivers to work with newer machines
oasys dsp card is stuck on win98 its a nice card and you can do some great things on it but only on a old mac or pc
not many synths on other dsp platforms
scope is not expensive at all i payed £10000 +£4000 +£3000 for an EII a macII and a soundtools and sd2 in 1985
now that was expensive and for 400k banks
resale value for that now is not much above a £1000
the cards are cheap in comparison and do much more than i could ever dream of
my first card i brought in 2000 and its still going strong 14 years later still supported with new soft and new plugs
my protools system which i paid £10000 for and a few 1000,s for interfaces became defunkt when they went from pci-x to pci-e and the upgrade path was pitance but if you did not upgrade then your system was none supported then they ditched it for native so i was glad i never upgraded
but with mac systems as soon as they changed the comp spec your old gear if you was upgrading was obsolete
but not the pci cards from scope i have upgraded several times and still my cards perform well
Re: Targeting Nividas CUDA as processing unit of the future?
WOW now Im really much more happy with my scope, than ever before. . . . you are right! Im with you.
out and about for music production. Are you still configguring your Studio music first!
Re: Targeting Nividas CUDA as processing unit of the future?
Also take into account 14 years ago when I bought my Pulsar, the 14DSP card costed around the same price as XITE-1 costs these days, with heaps less software included and there were no sales like today.
It's easy to be deceived by the fact so much functionality is now packed into a 1U box.
It's easy to be deceived by the fact so much functionality is now packed into a 1U box.
Re: Targeting Nividas CUDA as processing unit of the future?
ok, nothing goes faster to death like software. I have a dx7, running in a box called Hardware. . . . ., its running with software properiate, no support required for 30 years! , scope runs with a PC witch needs care like a baby everydays update. Somedays running better, next day sunndenly infected and ill, or killed by bluescreen, cause: even god doesnt know it, or maybe Jehova?
Thats why my musicmachine is not connected to the internet.
Never change a running system
Thats why my musicmachine is not connected to the internet.
Never change a running system
out and about for music production. Are you still configguring your Studio music first!
Re: Targeting Nividas CUDA as processing unit of the future?
Scope is software running on hardware - same as your DX7.anabella wrote:nothing goes faster to death like software
anabella wrote:scope runs with a PC witch needs care like a baby
anabella wrote:I have a 499€ PC
Bud Weiser wrote:You always pay what you get for.
Not because it is easy, but because it is hard...
yRe: Targeting Nividas CUDA as processing unit of the future
i have never had a blue screen or problems with crashing on my scope systems
if yours does then you have bad driver problems video cards or a rubbish motherboard or bad ram
my systems are rock solid with the 32bit os
if yours does then you have bad driver problems video cards or a rubbish motherboard or bad ram
my systems are rock solid with the 32bit os
Re: Targeting Nividas CUDA as processing unit of the future?
Never ever have a crash on my old OS9 mac with the cards.
Don't run anything else in parallel.
So Scope software is f* stable
Don't run anything else in parallel.
So Scope software is f* stable