|
Quick Question on DSP Cards (expreinced Producers only Please) (pg. 2)
|
View this Thread in Original format
| DJ-Igloo |
| thanks for the input guys, Pete my partner brought an Idea to me today about mabey getting a digital console insted of the UAD/Firewire etc cards what would be your view on that the console we would be looking at would be a Mackie Digital X bus or Mackie d8b board with the expansion cards, but I dont know if gong a console route would be better or worse then getting a powercore/UAD cards |
|
|
| thecYrus |
well, you can't compare a console/mixer with dsp card. it's something completly different.
but i wouldn't go back to the old days with external mixers and so. today everything happens inside the computer. it's just much more flexibel to have everything in the computer and have it saved in one single project files of your daw. total recall is the way to go.
and for dsp cards it only helps if you really "need" the plugins which are available only on those cards. today you can buy the whole sony oxford plugins as native vsts (you don't need a dsp card for them) so i don't really see a big point in those cards anymore.. |
|
|
| DigiNut |
| quote: | Originally posted by thecYrus
i think the time for those DSP cards is almost over. you can have native plugins which sound as good as those dsp stuff and you can run a lot more of those plugins on the CPU than on DSP. and with native plugins you are not limited to hardware (pci slots, ...) |
Here we go again with this silliness. You absolutely cannot perform more DSP tasks at one time on a general-purpose CPU than you can on a dedicated DSP unit (or several running in parallel, as the case is with these boards). Not even if the CPU is clocked 10 times faster. Not even if it's clocked 100 times faster.
To suggest otherwise flies in the face of all reason. I can think of no better proof than the Folding @Home project. Look at the statistics: 193444 active (windows) CPUs push out 184 TFLOPS (average: 0.95 MFLOPS per CPU), and just 958 active GPUs push out 57 TFLOPS (average: 59.5 MFLOPS/GPU). Considering that the average GPU is clocked at roughly 10% of the speed of the CPU in the same PC, that makes a typical GPU (roughly, again) 600 times more efficient than a general-purpose CPU.
And we aren't even looking specifically at DSP yet - just ordinary floating-point calculations in a GPU. DSP happens to be one area where certain specialized and complex functions which might ordinarily require a long string of FP operations can be parallelized in a single block and executed almost immediately. If a GPU can crunch the raw numbers 600 times faster than a CPU, a dedicated DSP can do the equivalent work another 10 times faster than the GPU.
I don't know how people can even think about saying that ordinary software plugins are just as good. Software certainly has its place, and the argument about DSP cards being proprietary and all the problems that entails is a valid one (albeit slightly misguided - look at the success of Reaktor), but software simply isn't a replacement for hardware performance-wise, no matter how many CPU cores your computer has or how fast they clock at.
I don't understand the "not limited to hardware" comment either - do you know something I don't? Have you figured out a way to run AUs on a Windows PC? |
|
|
| Existo22 |
| Word has it the univeral audio has better compressors and the tc powercore has better reverbs. |
|
|
| thecYrus |
| just to take it short: you can run a Sony Oxford EQ 11 times on a powecore mkII and you can run it 70+ on a core2duo e6600. it's exactly the same algorithmus so i doubt a powercore has anymore the power of a cpu. |
|
|
| r206 |
| quote: | Originally posted by DigiNut
Here we go again with this silliness. You absolutely cannot perform more DSP tasks at one time on a general-purpose CPU than you can on a dedicated DSP unit (or several running in parallel, as the case is with these boards). Not even if the CPU is clocked 10 times faster. Not even if it's clocked 100 times faster. |
Silliness? DSPs are very slow compared to todays GPUs and gp CPUs. For example the dsp56367 used in powercore mkII is clocked at 150MHz and can complete 1 instruction each cycle. Compare that to new GPUs like the r600 or multicore CPUs using SSE code. GPUs/CPUs are actually faster per clock cycle and they're clocked much faster.
| quote: | Originally posted by DigiNut
And we aren't even looking specifically at DSP yet - just ordinary floating-point calculations in a GPU. DSP happens to be one area where certain specialized and complex functions which might ordinarily require a long string of FP operations can be parallelized in a single block and executed almost immediately. If a GPU can crunch the raw numbers 600 times faster than a CPU, a dedicated DSP can do the equivalent work another 10 times faster than the GPU.
|
You don't use ordinary fp or int calculations for dsp? "certain specialized and complex functions" like? "require a long string of FP operations" ? mul-add? What are you talking about? "..can be parallelized in a single block and executed almost immediately" VLIW?
It's clear that you don't know anything about DSP so why write a long answer with incorrects "facts"?
I wouldn't buy a DSP card today. Use native plugins or buy a good hardware processor. |
|
|
| kitphillips |
Hmmm, well, seemse like we hijacked this thread:toothless
I reckon that these newer CPUs (Core 2 duos) can probably match a DSP quite well. As I understand it (and I'm not much of a computer nerd so I could be wrong) the problem with the old CPUs was that the instruction sets were never as comprehensive as the DSP cards were, hence the fact that a DSP could run faster, but the new SSE2 instruction set seems to have a good rep for audio processing, which may make it more evenly matched to DSP cards.
DSP will still have better instruction sets, but a new CPU blows them out of the water in terms of speed, multithreading, etc. Which means that a CPU will have a big edge most of the time now.
Back to the original debate: did you see that clip of ATB in his studio with his console and start drooling Igloo:p I did, but I think that those consoles are worth more than your average ferrari:eyes: , so I still say that the liquid mix is a better option. |
|
|
| flutlicht junky |
I think a hardware mixer could be a very good idea especially if you can use firewire to transfer the sound digitally to it for mixdown and to apply external hardware effects to it and then transfer the sound back.
It's the same kind of thing with using vsti versus hardware - its very intuitive to move a slide, turn a knob than look at a screen. And it might help you focus more on the sounds than what/how sound is being played. |
|
|
| mysticalninja |
| I thought if I got a UAD card all the UAD plugins would be running on the dsp card and my cpu wouldn't be affected at all? |
|
|
| DigiNut |
| quote: | Originally posted by thecYrus
just to take it short: you can run a Sony Oxford EQ 11 times on a powecore mkII and you can run it 70+ on a core2duo e6600. it's exactly the same algorithmus so i doubt a powercore has anymore the power of a cpu. |
That's because EQ is a ridiculously inefficient use of DSP resources. Just try doing the same thing with reverb or impulse processing or tube/speaker emulators.
The best analogy is storing a ton of 1-byte files on an NTFS-formatted hard disk with a cluster size of 32 KB. Storing 4096 of these files will not use 4 KB, it will actually eat up 128 MB because of the minimum allocation unit.
I could therefore say, and this would be identical to your logic, that a 16 GB FAT16-formatted drive with a 2 KB cluster size could "hold much more data" than a 160 GB NTFS drive with a 32 KB cluster size. This is technically true, but only in the farcical isolated case of storing nothing but a bunch of tiny files with a few bytes up to 1 KB each. Clearly a 160 GB drive can hold a lot more than a 16 GB drive; you can't store a 20 GB database on the 16 GB drive!
| quote: | Originally posted by r206
Silliness? DSPs are very slow compared to todays GPUs and gp CPUs. For example the dsp56367 used in powercore mkII is clocked at 150MHz and can complete 1 instruction each cycle. Compare that to new GPUs like the r600 or multicore CPUs using SSE code. GPUs/CPUs are actually faster per clock cycle and they're clocked much faster. |
I explained in the above post why the clock speed is hardly relevant. Anyway, the best GPU you can buy today is clocked at 450 MHz, so a DSP clocked at 150 MHz is hardly much of a downgrade in that respect. Most DSPs also have internal instruction clocks at 200-400 MHz; it's just the DSP clock that's slower. Also, the "slow" clock speed only applies to PowerCore; I believe UAD DSPs are clocked at 400 MHz.
| quote: | You don't use ordinary fp or int calculations for dsp? "certain specialized and complex functions" like? "require a long string of FP operations" ? mul-add? What are you talking about? "..can be parallelized in a single block and executed almost immediately" VLIW?
It's clear that you don't know anything about DSP so why write a long answer with incorrects "facts"? |
No, it's clear that you don't know what you're talking about because you've posted two examples ("mul-add" and "VLIW", by the latter I assume you're talking about Very Long Instruction Word) which have very little to do with DSP, other than the fact that a couple of DSPs sorta-kinda look like a VLIW architecture these days. I studied electrical engineering with a significant amount of that being in signal processing, so yes, I do know a few things about it.
Some of the instructions I'm referring to are:
- Dot Product
- Complex arithmetic (as in complex/imaginary numbers, not "difficult")
- FIRS (Symmetric Finite Impulse Response)
- MAC (Multiply-Accumulate - this is your "mul-add")
- Polynomial expansion/factoring instructions
- Least-mean-square (LMS) calculations
- Vector distance or square-distance, with or without ABS
That's by no means an exhaustive list, and yes, GPUs do implement some of these but not all of them. Some of them are generic (such as vector distance) but some are definitely DSP specific (i.e. FIRS, which is one of the most common operations in signal processing). Typically these take 1 clock cycle on a DSP chip; sometimes 2 or 3.
Many of these are classified as SIMD (Single Instruction, Multiple Data) operations, although I assume you must have known that given the level of expertise you're implying. Yes, general-purpose CPUs with add-ons like MMX do implement some SIMD operations, but unsurprisingly, they are general-purpose ones that don't really help DSP or graphics-intensive applications. If they did, we wouldn't need fancy GPUs.
Bottom line: you can (and do) waste hundreds or thousands of clock cycles performing the same operations that take one or maybe a couple of clock cycles on a DSP. Also, DSP is almost infinitely parallelizable; you can stick 4 processors or 16 processors on the same interface for minimal cost. I hear people talking about the wonders of threading and dual-core CPUs, but I doubt that any of you, unless you are programmers (which I am, in addition to the elec. eng. background) really know what the hell you're talking about. Multi-threading on a PC involves context switches and task scheduling, which are hideously expensive. Most DSP units come with at least 4 DSP cores and because of the way DSP works (you gradually use up resources and once you're out, you're out), there is no need for context switches and hence, no overhead whatsoever.
So, you ask, why doesn't every plugin rape the CPU of most modern computers? Duh - it's because the software algorithms take a ton of shortcuts. That's why software reverbs still suck, but really simple ones like EQ are fine. Beefy plugins like convolutions DO rape the CPU, as do most heavy-duty analog modelling synths like z3ta+ or Reaktor and some of the instruments people make on it.
Please people, for 's sake, stop trying to analyze this and just accept the fact that General-Purpose CPUs can't match the throughput of a GPU or DSP. That doesn't mean you need a DSP, nor does it mean that software sucks, it just means that they are two totally different things and neither is a good replacement for the other. Can we end this stupid debate now? |
|
|
| Mr Rogers |
| quote: | Originally posted by DigiNut
That's because EQ is a ridiculously inefficient use of DSP resources. Just try doing the same thing with reverb or impulse processing or tube/speaker emulators.
The best analogy is storing a ton of 1-byte files on an NTFS-formatted hard disk with a cluster size of 32 KB. Storing 4096 of these files will not use 4 KB, it will actually eat up 128 MB because of the minimum allocation unit.
I could therefore say, and this would be identical to your logic, that a 16 GB FAT16-formatted drive with a 2 KB cluster size could "hold much more data" than a 160 GB NTFS drive with a 32 KB cluster size. This is technically true, but only in the farcical isolated case of storing nothing but a bunch of tiny files with a few bytes up to 1 KB each. Clearly a 160 GB drive can hold a lot more than a 16 GB drive; you can't store a 20 GB database on the 16 GB drive!
I explained in the above post why the clock speed is hardly relevant. Anyway, the best GPU you can buy today is clocked at 450 MHz, so a DSP clocked at 150 MHz is hardly much of a downgrade in that respect. Most DSPs also have internal instruction clocks at 200-400 MHz; it's just the DSP clock that's slower. Also, the "slow" clock speed only applies to PowerCore; I believe UAD DSPs are clocked at 400 MHz.
No, it's clear that you don't know what you're talking about because you've posted two examples ("mul-add" and "VLIW", by the latter I assume you're talking about Very Long Instruction Word) which have very little to do with DSP, other than the fact that a couple of DSPs sorta-kinda look like a VLIW architecture these days. I studied electrical engineering with a significant amount of that being in signal processing, so yes, I do know a few things about it.
Some of the instructions I'm referring to are:
- Dot Product
- Complex arithmetic (as in complex/imaginary numbers, not "difficult")
- FIRS (Symmetric Finite Impulse Response)
- MAC (Multiply-Accumulate - this is your "mul-add")
- Polynomial expansion/factoring instructions
- Least-mean-square (LMS) calculations
- Vector distance or square-distance, with or without ABS
That's by no means an exhaustive list, and yes, GPUs do implement some of these but not all of them. Some of them are generic (such as vector distance) but some are definitely DSP specific (i.e. FIRS, which is one of the most common operations in signal processing). Typically these take 1 clock cycle on a DSP chip; sometimes 2 or 3.
Many of these are classified as SIMD (Single Instruction, Multiple Data) operations, although I assume you must have known that given the level of expertise you're implying. Yes, general-purpose CPUs with add-ons like MMX do implement some SIMD operations, but unsurprisingly, they are general-purpose ones that don't really help DSP or graphics-intensive applications. If they did, we wouldn't need fancy GPUs.
Bottom line: you can (and do) waste hundreds or thousands of clock cycles performing the same operations that take one or maybe a couple of clock cycles on a DSP. Also, DSP is almost infinitely parallelizable; you can stick 4 processors or 16 processors on the same interface for minimal cost. I hear people talking about the wonders of threading and dual-core CPUs, but I doubt that any of you, unless you are programmers (which I am, in addition to the elec. eng. background) really know what the hell you're talking about. Multi-threading on a PC involves context switches and task scheduling, which are hideously expensive. Most DSP units come with at least 4 DSP cores and because of the way DSP works (you gradually use up resources and once you're out, you're out), there is no need for context switches and hence, no overhead whatsoever.
So, you ask, why doesn't every plugin rape the CPU of most modern computers? Duh - it's because the software algorithms take a ton of shortcuts. That's why software reverbs still suck, but really simple ones like EQ are fine. Beefy plugins like convolutions DO rape the CPU, as do most heavy-duty analog modelling synths like z3ta+ or Reaktor and some of the instruments people make on it.
Please people, for 's sake, stop trying to analyze this and just accept the fact that General-Purpose CPUs can't match the throughput of a GPU or DSP. That doesn't mean you need a DSP, nor does it mean that software sucks, it just means that they are two totally different things and neither is a good replacement for the other. Can we end this stupid debate now? |
i was gonna say the exact same thing |
|
|
|
|
