Multiple graphics cards could lead to unimaginable computer power provided that software companies implement them.
http://www.wired.com/techbiz/it/news/2008/06/gpu_power
Your Own Super Computer
The idea of using a graphics card for computing is not new but having multiple cards takes it to a new level. One thing for sure, disposing of computers is wasteful. If there was a way to simply buy another card and greatly increase performance, it would benefit consumers and be less destructive to the environment. It's not clear if these multiple cards will be used for greater computer power but I hope so. It sure beats buying a whole new computer every time you want more power. There has to be a better way than what we have.
It would be nice if the good folks at Sonic Core could focus on software and let people who really know what they are doing make the PCIe cards. Sadly, it's just easier to continue with the tired old ways than to learn how to do things differently. Graphics cards become dirt cheap after just a couple of years, not so for Scope cards.
Braincell, you may have missed the discussion here about NVidea's free SDK - roughly a year back or so...
This is certainly a great tool to program your private weather forecast or anything else based on massive amounts of static data, but it is completely useless in the real-time domain.
You see, people prefere ASIO buffersizes of 64 bytes and less for the sake of very small latencies - but these GPUs live up to their 'performance specs' only with buffers thousands of times larger.
This is not a 'general' processing power, but a highly application specific one.
Supercomputer performance is only achieved with problems that consist of a huge amount of elements that all need the same processing applied, that's why they excel in bitmap processing.
There's an Italien company (the name is so stupid that I forget it time and again, sorry) who had announced convolution based audio processing with these cards, or more precisely an adaption of their native product.
This is no news at all, as is the cascading of GPUs which hardcore gamers do for a while.
The core problem remains the same and it's been here for decades - you need an 'engine' that is capable to analyze processing advices (let's assume a pseudocode source) and detect structures that are 'parallelizable, so it can transform them into instructions for such devices.
The current solution for this processor is you, the programmer.
First you have to think about the process you'd like to handle
Then you'll have to find a matching 'method pattern' in the world of graphics processing and you can start.
The NVidea SDK intends to ease this process by providing a more abstract representation (and terminolgy), but basically it (still?) suffers from those conditions.
NVidea kind of declared it as '...a tool to write future tools...' and not an application developing environment.
I've bought a GForce 8600 to have a peek at the technology, but found it way too inefficient to start digging deeper, as I don't have a specific problem at hand that could really benefit.
There is at least one other member here who successfully used the technology for scientific (statistical?) data, according to his post.
cheers, Tom
This is certainly a great tool to program your private weather forecast or anything else based on massive amounts of static data, but it is completely useless in the real-time domain.
You see, people prefere ASIO buffersizes of 64 bytes and less for the sake of very small latencies - but these GPUs live up to their 'performance specs' only with buffers thousands of times larger.
This is not a 'general' processing power, but a highly application specific one.
Supercomputer performance is only achieved with problems that consist of a huge amount of elements that all need the same processing applied, that's why they excel in bitmap processing.
There's an Italien company (the name is so stupid that I forget it time and again, sorry) who had announced convolution based audio processing with these cards, or more precisely an adaption of their native product.
This is no news at all, as is the cascading of GPUs which hardcore gamers do for a while.
The core problem remains the same and it's been here for decades - you need an 'engine' that is capable to analyze processing advices (let's assume a pseudocode source) and detect structures that are 'parallelizable, so it can transform them into instructions for such devices.
The current solution for this processor is you, the programmer.
First you have to think about the process you'd like to handle
Then you'll have to find a matching 'method pattern' in the world of graphics processing and you can start.
The NVidea SDK intends to ease this process by providing a more abstract representation (and terminolgy), but basically it (still?) suffers from those conditions.
NVidea kind of declared it as '...a tool to write future tools...' and not an application developing environment.
I've bought a GForce 8600 to have a peek at the technology, but found it way too inefficient to start digging deeper, as I don't have a specific problem at hand that could really benefit.
There is at least one other member here who successfully used the technology for scientific (statistical?) data, according to his post.
cheers, Tom
I'm sorry but this is nothing but a polemic statement by someone not knowing what he's talking about.braincell wrote:It would be nice if the good folks at Sonic Core could focus on software and let people who really know what they are doing make the PCIe cards. Sadly, it's just easier to continue with the tired old ways than to learn how to do things differently.
Make up your mind about massively parallel architectures on one hand and realtime DSP processing on the other.
You will find that the SonicCore approach at least scales way better than the general purpose multicore rubbish by your OS provider
not after a couple of years, Braincell - after a couple of millions in sales quantity.Graphics cards become dirt cheap after just a couple of years, not so for Scope cards.
cheers, Tom
Multicore processors are not taken advantage of very frequently by software developers. As I understand it, Cubase does split the processors and use one for effects. I do not have a multicore CPU yet.
It would be nice if you could plug in more CPU modules the way you can plug in more graphics cards. The original intent of Creamware as it was marketed was that the DSP chips take the load of the CPU. As computers become more powerful, this distinction will be totally pointless and a waste of money.
It would be nice if you could plug in more CPU modules the way you can plug in more graphics cards. The original intent of Creamware as it was marketed was that the DSP chips take the load of the CPU. As computers become more powerful, this distinction will be totally pointless and a waste of money.
astroman wrote: You will find that the SonicCore approach at least scales way better than the general purpose multicore rubbish by your OS providernot after a couple of years, Braincell - after a couple of millions in sales quantity.Graphics cards become dirt cheap after just a couple of years, not so for Scope cards.
cheers, Tom
Your OS provider is better. Too bad few people want to buy them!
-
kensuguro
- Posts: 4434
- Joined: Sun Jul 08, 2001 4:00 pm
- Location: BPM 60 to somewhere around 150
- Contact:
i do agree that this would be really cool if it were to come true... There are types of processes that benefit greatly from GPU type calculations, like running simulations and stuff.. I once saw a liquid simulater that was written to run off GPU and it performed extremely well. I think it may have been one of the nvidia demos.
I'm not sure if it has any use in terms of every day computing, but it would be cool to have some sort of card that you can jam in all the old cpus you have, and just add it to your new machine. Probably not that powerful, but atleast it's a collective effort and like braincell says, it's a very efficient and green (or unwasteful, which ever you like) solution. Kind of like reviving an old machine with linux. Works very well for poor third world countries, esp schools and stuff I think. And here we are, buying new macbooks and ipods like candy.
I'm not sure if it has any use in terms of every day computing, but it would be cool to have some sort of card that you can jam in all the old cpus you have, and just add it to your new machine. Probably not that powerful, but atleast it's a collective effort and like braincell says, it's a very efficient and green (or unwasteful, which ever you like) solution. Kind of like reviving an old machine with linux. Works very well for poor third world countries, esp schools and stuff I think. And here we are, buying new macbooks and ipods like candy.
Braincell as you like to split my quotes beyond context, I will simply repeat it with other words:
you diss Creamware/SonicCore to be old-fashioned - may I assume you mean behind in technology ?
yet the distribution of procesing power over DSPs works several times more effective than those multi-core optimizations of any (PC) OS manufacturer or those by developers of native software, like Steinberg, as you mentioned yourself.
Add twice the amounts of DSPs - you'll get twice the amount of voices
add a Quadcore and it will bring you 50% ahead at best, using only 25-30% of it's supposed resources.
You will exceed this rate only in hand-coded parts of the program.
And now try to imagine 'hand-coded' in a context where programs have millions of lines of source code...
If you make yourself a little bit more familiar with the conditions of realtime processing, you will also learn that a general purpose CPU is limited already by it's architecture, using less than 20% of the available 'power' for effective processing.
Most of the time it spends idle, waiting on data being aligned for processing.
this 'processor power thing' hyped in press is a fairy tale, not much more than marketing bullshit. It relates to abstract figures and the benchmarks aren't called 'synthetic' for nothing.
when multitrack recording started to gain popularity there was a dude from Las Vegas who wrote the S.A.W Studio application which could do (say) 24 tracks of live recording on a Pentium 200 (or so).
He showed the program on NAMM and people said there must be a cheat, as it would be impossible (according to their experience in Windows)
In fact it would have been impossible, and the clue was that this dude bypassed all and everything related to Windows audio, handcoding all the basic IO and file operations.
So what's the point in environementalism then ?
Wanna sue Micro$oft for wasting CPU power or make them pay a dedicated tax ?
Btw a GPU is not even remotely a solution in this context, as it cannot be programmed to solve general purpose programs effectively. If it processes (say) 4 elements, 252 'lanes' will be idle, but still consume power.
For curiosity I've downclocked my GPU for over 50% - it didn't change much of heat emission.
cheers, Tom
you diss Creamware/SonicCore to be old-fashioned - may I assume you mean behind in technology ?
yet the distribution of procesing power over DSPs works several times more effective than those multi-core optimizations of any (PC) OS manufacturer or those by developers of native software, like Steinberg, as you mentioned yourself.
Add twice the amounts of DSPs - you'll get twice the amount of voices
add a Quadcore and it will bring you 50% ahead at best, using only 25-30% of it's supposed resources.
You will exceed this rate only in hand-coded parts of the program.
And now try to imagine 'hand-coded' in a context where programs have millions of lines of source code...
If you make yourself a little bit more familiar with the conditions of realtime processing, you will also learn that a general purpose CPU is limited already by it's architecture, using less than 20% of the available 'power' for effective processing.
Most of the time it spends idle, waiting on data being aligned for processing.
this 'processor power thing' hyped in press is a fairy tale, not much more than marketing bullshit. It relates to abstract figures and the benchmarks aren't called 'synthetic' for nothing.
when multitrack recording started to gain popularity there was a dude from Las Vegas who wrote the S.A.W Studio application which could do (say) 24 tracks of live recording on a Pentium 200 (or so).
He showed the program on NAMM and people said there must be a cheat, as it would be impossible (according to their experience in Windows)
In fact it would have been impossible, and the clue was that this dude bypassed all and everything related to Windows audio, handcoding all the basic IO and file operations.
So what's the point in environementalism then ?
Wanna sue Micro$oft for wasting CPU power or make them pay a dedicated tax ?
Btw a GPU is not even remotely a solution in this context, as it cannot be programmed to solve general purpose programs effectively. If it processes (say) 4 elements, 252 'lanes' will be idle, but still consume power.
For curiosity I've downclocked my GPU for over 50% - it didn't change much of heat emission.
cheers, Tom
There is plenty of CPU power these days but not enough PCI bandwidth. CPU power is the current state of the art in music applications *not* dedicated DSP chips. It's nice when you buy a new computer to see an instant increase in performance, however you won't have better performance in your Scope cards because they are independent. What was once an advantage, now becomes a liability. I guess Sonic Core's answer to this is the xite, an interesting solution at an ungodly price.
The problem with using gpu chips for processing is the huge latency.See uad1.
And the uad cards sell like chicken....they use a now 10 year old long discontinued gpu.And scope is a very very low latency environment with sharcs...and these chips get used a LOT by other highend manufacturers like euphonics aswell.
And the uad cards sell like chicken....they use a now 10 year old long discontinued gpu.And scope is a very very low latency environment with sharcs...and these chips get used a LOT by other highend manufacturers like euphonics aswell.
Actually NVidia had purchased this little company which halted all of the great prgramming ideas they had.
I have heard nothing since a few years ago about this, but I was interested in this for Gigastudio 2.54, But then GS3 came out w/ Gigapulse and I bought some Scope DAW's.
It made sense to put CPU killing effects processes on the GPU, at least back then it did.
http://www.nforcershq.com/article1853.html
I have heard nothing since a few years ago about this, but I was interested in this for Gigastudio 2.54, But then GS3 came out w/ Gigapulse and I bought some Scope DAW's.
It made sense to put CPU killing effects processes on the GPU, at least back then it did.
http://www.nforcershq.com/article1853.html
-
siriusbliss
- Posts: 3118
- Joined: Fri Apr 06, 2001 4:00 pm
- Location: Cupertino, California US
- Contact:
3 Scope cards and 4 monitors here - going back 10 years now with two of my original cards.braincell wrote:I just read these multiple cards saturate the PCI bus... a problem all Scope users are all too familiar with! It would be pretty cool to have 6 monitors though!
10 years ago there WAS problems with the PCI-bus 'saturation' (sharing IRQ's, etc'), but now nowadays.
GPU's are best for math-coprocessing graphics intensive apps., but not really necessary for streaming audio through the host application IMO.
Greg
Xite rig - ADK laptop - i7 975 3.33 GHz Quad w/HT 8meg cache /MDR3-4G/1066SODIMM / VD-GGTX280M nVidia GeForce GTX 280M w/1GB DDR3
I can not use my firewire hard drive with the Ambience reverb module using the latest Scope cards. This *is* a problem period. Ambience is my favorite module and the Firewire hard drive is my favorite drive. Your experience may differ but to me it totally sucks. I get the PCI overflow message the second I start the module. You don't have the same hardware as I do so I can not say what you wrote is a lie but it is wrong and misinformed. Just to borrow a phrase from my friend Astroman:
"I'm sorry but this is nothing but a polemic statement by someone not knowing what he's talking about. "
"I'm sorry but this is nothing but a polemic statement by someone not knowing what he's talking about. "
Pci or pci-e firewire card? Onboard fw? Agere/ti/via chipset? Info? Help? Ranting again? ??? ?? ? Wtf do you want, need or trying to archieve??? Questionmark!braincell wrote:I can not use my firewire hard drive with the Ambience reverb module using the latest Scope cards. This *is* a problem period. Ambience is my favorite module and the Firewire hard drive is my favorite drive. Your experience may differ but to me it totally sucks. I get the PCI overflow message the second I start the module. You don't have the same hardware as I do so I can not say what you wrote is a lie but it is wrong and misinformed. Just to borrow a phrase from my friend Astroman:
"I'm sorry but this is nothing but a polemic statement by someone not knowing what he's talking about. "
-
siriusbliss
- Posts: 3118
- Joined: Fri Apr 06, 2001 4:00 pm
- Location: Cupertino, California US
- Contact:
So, your firewire is running through your PCI bus? Well, ok, then you have a problem there don't you.braincell wrote:I can not use my firewire hard drive with the Ambience reverb module using the latest Scope cards. This *is* a problem period. Ambience is my favorite module and the Firewire hard drive is my favorite drive. Your experience may differ but to me it totally sucks. I get the PCI overflow message the second I start the module. You don't have the same hardware as I do so I can not say what you wrote is a lie but it is wrong and misinformed. Just to borrow a phrase from my friend Astroman:
"I'm sorry but this is nothing but a polemic statement by someone not knowing what he's talking about. "
Not misinformed here. Running plenty of modules off firewire with no issues, including ROMplers, sampler libraries, etc.
Check to see if your firewire is sharing IRQs with Scope hardware. It shouldn't, but who knows. Otherwise what happens when you run it off local harddrive?
Greg
Xite rig - ADK laptop - i7 975 3.33 GHz Quad w/HT 8meg cache /MDR3-4G/1066SODIMM / VD-GGTX280M nVidia GeForce GTX 280M w/1GB DDR3
It's ok as long as I don't use the firewire and Ambience. All peripherals connect through the PCI bus ultimately. I can not change the IRQ settings on this computer.
VSTi never have any such problem. Only Scope cards have this total meltdown. It's frustrating let me tell you. I have been going through this for years. The older computer with the new cards works better. I can use the Ambience on that. I haven't tried the firewire on that one because it doesn't have a firewire port. This computer is faster and the firewire drive is faster but it is not a good computer for Scope. This is hell I tell you! I want to plug in my hardware and have it work period. I have 4 Creamware and Scope cards. I don't need that many but I had to buy extra cards to replace the type I cards because of PCI overflow problems. This is a nightmare on and off for a decade.
This is how my old computer is configured. It took a really long time to figure this out and it is the only way it would work:
Slot 1 IRQ5
Slot 2 Auto
Slot 3,4 IRQ 5
Slot 5,6 Auto
VSTi never have any such problem. Only Scope cards have this total meltdown. It's frustrating let me tell you. I have been going through this for years. The older computer with the new cards works better. I can use the Ambience on that. I haven't tried the firewire on that one because it doesn't have a firewire port. This computer is faster and the firewire drive is faster but it is not a good computer for Scope. This is hell I tell you! I want to plug in my hardware and have it work period. I have 4 Creamware and Scope cards. I don't need that many but I had to buy extra cards to replace the type I cards because of PCI overflow problems. This is a nightmare on and off for a decade.
This is how my old computer is configured. It took a really long time to figure this out and it is the only way it would work:
Slot 1 IRQ5
Slot 2 Auto
Slot 3,4 IRQ 5
Slot 5,6 Auto