NVIDIA goes VST?

astroman · Post by **astroman** » Fri Sep 10, 2004 4:53 am

On 2004-09-10 02:46, symbiote wrote:
... POVRay uses more than 15%, also compiling stuff takes more than 15%,...

you misunderstood the meaning of the '15% figure' - it's not about CPU load, but about CPU potential if the code would be more optimized .

I dunno POV in detail, but from all I've read about it I'm rather convinced that it's far from highly optimized graphic routines (and assembly wizzardry)

That kind of coding is not really my cup of tea, but I've learned programming in a time when it was absolutely essential that you found ways to make your code perform faster.

CPUs were clocked from 8-30 MHZ those years and I still occasionally use apps from this time which easily outperform M$'s crap on a CPU running at least 30 times faster.

So the estimation that a routine can be optimized for a factor of 50 isn't totally absurd

You'll find the same information in any 'classic' book about programming algorithms and strategies.

The current public opinion about this subject is something like 'well we have fast CPUs, so there's no need to fiddle around - we don't have time for this...'

As wrong as can be - there's a physical limit in CPU clocking and it's almost reached.
Or are you seriously expecting a 50 GIG P4 anywhere in near future ?

On 2004-09-10 03:47, R-type wrote:
It's a sweet idea, if I was Creamware I would be looking at recompiling their awesome range of plugins to a new platform. ...

do you have any idea what it takes to transfer a product consisting of a 150 man years developement effort to a new platform ?

And in case you didn't notice: the math core routines aren't by Creamware, but by Analog devices...

You may add another 50 man years to transfer (better re-engineer) that part of the system.
But while it may be possibly to hire some clever dudes for the GUI stuff, you will just fail on the latter issue due to the lack of people capable to deal with that kind of coding.

cheers, Tom

[ This Message was edited by: astroman on 2004-09-10 05:59 ]

cleanbluesky · Post by **cleanbluesky** » Fri Sep 10, 2004 7:32 am

I think you are right Tom, todays processors are REALLY fast compared to what has come before and I think all that extra power has made programmers lazy. And lets face it, whenever power is called into question the general rule is to upgrade your computer, not to expect programmers to optimise...
A good example of enhanced coding can be seen when some game manufacturers or GPU card makers release software updates. The same games can sometimes suddenly run faster or smoother...

symbiote · Post by **symbiote** » Fri Sep 10, 2004 9:56 am

On 2004-09-10 05:53, astroman wrote:
you misunderstood the meaning of the '15% figure' - it's not about CPU load, but about CPU potential if the code would be more optimized .

I dunno POV in detail, but from all I've read about it I'm rather convinced that it's far from highly optimized graphic routines (and assembly wizzardry)
That kind of coding is not really my cup of tea, but I've learned programming in a time when it was absolutely essential that you found ways to make your code perform faster.

CPUs were clocked from 8-30 MHZ those years and I still occasionally use apps from this time which easily outperform M$'s crap on a CPU running at least 30 times faster.

So the estimation that a routine can be optimized for a factor of 50 isn't totally absurd
You'll find the same information in any 'classic' book about programming algorithms and strategies.

The current public opinion about this subject is something like 'well we have fast CPUs, so there's no need to fiddle around - we don't have time for this...'

As wrong as can be - there's a physical limit in CPU clocking and it's almost reached.
Or are you seriously expecting a 50 GIG P4 anywhere in near future ?

While I agree that 3GHz is a bit much for text processing and pushing emails around, saying that all algorithms around are invariably unoptimized is a bit much. True for Word and things, alot less true for other stuff. (Not to mention that *no one* in their right (or wrong) mind would want to mess around with assembly optimizations on current Intel CPUs -- icky, iiiick.)

About POVRAY, it's been around for, what, 10 years? At least 8, so you can count easily 80-100 man-years of work invested in it. I'm sure it's pretty nicely optimized in alot of aspects, it was started on those 8-30MHz machines you are talking about. Assembly is all nice and wonderful and I love it too, but assembly can't automagically make some mathematically-intense processes more simple.

Also, I've re-read your original post, and it sounds like you've never looked at a convolution mathematically. Saying it can't be done fast because programmers are lazy is a bit off-target. It's a pretty well studied problem/process, and alot of people would gain from finding a faster way to implement it. Making it faster is as much a matter of choosing the right processing architecture than a matter of programming.

Besides it's not that programmers are lazy (some are, some others are still obsessed by optimization,) it's that the big players in the field profit alot from continuous hardware and software upgrades. If they'd release a hugely stable machine that did everything and never needed to be upgraded, everyone would stop buying computers every 2 years, and they'd lose the massive cash input. And since they are sitting on $50 billion, it's pretty easy for them to get their way

Saying anything could be further optimized is pretty easy, doing it is something else entirely.

BTW, about current GPUs, the problem is that getting back the data from the card is really slow. So it does 40 GFLOPs, but you really can't get back any of that information easily, hence it's not really useful yet. Maybe someone can hack something with DVI to get the stream back into the computer, or convince video card manufacturers to add a data path back to the computer that's fast enough for the job.

Anyway, I'd really like to see a 8-30MHz machine raytracing 292k spheres with anti-aliasing and radiosity

Maybe with a bit of liquid nitro!

astroman · Post by **astroman** » Fri Sep 10, 2004 4:21 pm

On 2004-09-10 10:56, symbiote wrote:
... Also, I've re-read your original post, and it sounds like you've never looked at a convolution mathematically. Saying it can't be done fast because programmers are lazy is a bit off-target. ...

if you refer to this statement:
It could be programmed much better if people had the time (and expertise)

this isn't exactly calling someone lazy.
It's about rushing things out too early to get a foot in the market

I've also quoted a certain 'inability' of programmers, which again isn't related to lazyness, but that some techniques are (for obvious reasons) out of focus today.
Imho most programmers seem to ignore a lot of possibilities - be it conciously or not.

Of course programmers are almost forced to stick with certain libraries (say M$ VisualCrapp++), and that's (one of) the reason(s) why I'm definetely out of this business...

I HAVE worked with an excellent lib once and have a good idea how such a thing has to be organised

Since I'm only in it for the money

I'm more concerned with database stuff than with low level bit-shifting and register switching - but I admit to have had quite a laugh when reading about the 'next big thing' in M$'s programming strategies, the F sharp meta stuff (Redmonds comment: we have to get things more straight...).

I am using a similiar system with some interuptions for as long as back to 1990 - finally M$ 'admitted' the stuff they preached for the last decade just wasn't it...

But you're absolutely right that I never cared much about the math of convolution - as mentioned, it's not my cup of tea and I admit I'm a comlplete ignorant when it comes to math theory.
But I'm pretty good in finding solutions and in putting together the pieces of the puzzle

I know convolution from graphics programming (in the context of aquiring and archiving medical ultrasound images) and I'm aware that one can extend the subject pretty far beyond the current reverb hype.

Nevertheless it's just one part of audio processing (and fits particularily good with the architecture of a GPU board), but let's not forget the (majority) of other algorithms.

cheers, Tom

rezzer · Post by **rezzer** » Sat Sep 11, 2004 12:18 am

This idea has been bouncing around for a while, these guys are the first to actually announce a product that uses the GPU to accelerate audio processing that I've heard of though.

Using the GPU for processing other things isn't really tied to a specific manufacturer, it's something that has become a viable option with the advent of pixel shader languages. These languages (there are a couple versions now) are sort of C like in appearance but are designed to work closely with the hardware. There are a couple of research projects going on right now where people are trying to make an even more general language that can be run on the GPU instead of basically hacking the shader languages to do something other than shading things.

You can read more about one of these research projects at http://www.gpgpu.org/

There's another similar project called BrookGPU or something, I don't have a link handy though.

Ultimately we can expect the GPU on our display cards to be used for pretty much any heavy duty parallel processing friendly task (such as audio signal processing). Moreover, we can expect it to be commonplace since the next version of Windows coming down the pike will require a display card that supports the shader languages needed to do this stuff. Once developers can depend on something being available, like a keyboard, mouse or sometime soon a GPU, they tend to use them whenever it's appropriate.

[ This Message was edited by: rezzer on 2004-09-11 01:19 ]

valis · Post by **valis** » Sat Sep 11, 2004 4:22 am

UAD-1 is was a gpu initially so technically it predates this (I believe that was stated in this thread but I thought I'd restate it

).

AndreD · Post by **AndreD** » Sat Sep 11, 2004 5:37 am

But doas anybody knows what kind of GPU?
That was a big secret some years ago..

astroman · Post by **astroman** » Sat Sep 11, 2004 7:35 am

I bet it's a design based on a (totally obsolete for some time) NEC Power VR accelerator - 50 bucks for planetZ if I'm wrong...

cheers, Tom

cleanbluesky · Post by **cleanbluesky** » Sat Sep 11, 2004 8:49 am

Even if the UAD-1 is a GPU, I think the most important thing is being able to tap into a hardware market that is constantly being developed and has a lot of money behind it, which means that newer models will be constantly brought out. In such a market, we would not be sitting with anything older than 6 months in our computer and still calling it top-of-the-line.

valis · Post by **valis** » Sat Sep 11, 2004 8:54 am

Why is it so important to be within 6 months of the top of the line?

Post by **garyb** » Sat Sep 11, 2004 12:53 pm

damn, most of the hardware that is coveted by commercial studios is at least 20 years old.......

Post by **hubird** » Sat Sep 11, 2004 8:32 pm

On 2004-09-11 09:54, valis wrote:
Why is it so important to be within 6 months of the top of the line?

agree, it's even risky for the workflow, you prob have to face soft- and/or hardware conflicts, and all the configuring actually makes you forget making music

wayne · Post by **wayne** » Sat Sep 11, 2004 8:46 pm

i may be fooling myself, but i am making professional work on a PIII 800

Post by **hubird** » Sat Sep 11, 2004 9:21 pm

and me on a (mac G4) 733 MHz... it provided me all songs on http://www.ezsound.nl

(the only thing I regret is CWA prohibits me of changing to OS-X, which is the better (more stable) system)

Be smart, think art

[ This Message was edited by: hubird on 2004-09-11 22:26 ]

wayne · Post by **wayne** » Sat Sep 11, 2004 9:25 pm

and we're both on vst32 still

Post by **hubird** » Sat Sep 11, 2004 9:27 pm

cleanbluesky · Post by **cleanbluesky** » Sun Sep 12, 2004 5:16 am

On 2004-09-11 13:53, garyb wrote:
damn, most of the hardware that is coveted by commercial studios is at least 20 years old.......

Yah, thats very true, but I doubt that many studios are coveting digital products that are 20+ years old. You don't need super-power to make good music, but being able to acess it if you want it is a huge advantage. Graphics cards are always being made faster to compete, this is a huge advantage if you want to be unrestricted when using processing-heavy FX.

Post by **garyb** » Sun Sep 12, 2004 5:28 am

without lots of money, you're bound to feel a little restricted, otherwise you'd just get more hardware and computers....

*edit* and you should hear them pine for their old ataris and amigas......

[ This Message was edited by: garyb on 2004-09-12 06:57 ]

valis · Post by **valis** » Sun Sep 12, 2004 6:07 am

Music making as a commodities trading game...

astroman · Post by **astroman** » Sun Sep 12, 2004 6:15 am

well, I'm admittedly a non-gamer, so I'd much appreciate if some of the 'unused' power of the GPU could be applied to some other processing

On the other hand I do have the suspect, that a lot of that processing power (aside from optimized data paths) simply exists because that thing can operate relatively undisturbed by the OS...

A GPU is still clocked a 10 magnitude lower than the main CPU

As mentioned there's only this one product released yet, while the idea itself is known for quite some time.

It's probably not that easy to just transfer existing audio processing code - and in fact I'd more expect completely new strategies to yield really high performance results.
But that's far more demanding even at the design stage and will take a lot of time.

Imho the potential power of whatever processing unit is of high marketing, but few practical use

cheers, Tom