When to expect ps 3.0 hardware?

[Nick]

I was wondering when the first hardware with support for ps 3.0 is going to hit the market...

The only reason I would like to know this is because I'm adding support for it in my software renderer and I want to know how much I have to rush it. It would be the first implementation at interactive framerates (100 times faster than REF) so game developers could really test it. Most games take several years to be developed and they target future hardware. But if hardware support is around the corner then there's no need to sweat over it (in that case please give me DirectX 10 specifications).

So do I have weeks, months or years? Give me your best guess.

 

[Ailuros]

Give me your best guess.

Spring 04'

 

[Arun]

Launch: Comdex 2003 - NV40
Limited avaibility: Christmas 2003 or January 2004
Real avaibility: February or March 2004

That is, unless nVidia fails to deliver - again.
The R420 will most likely launch slightly later, in early 2004.

Uttar

 

[andypski]

The only reason I would like to know this is because I'm adding support for it in my software renderer and I want to know how much I have to rush it. It would be the first implementation at interactive framerates (100 times faster than REF) so game developers could really test it.

I hope it's more than 100 times faster than REF if you want to call it interactive - I've sat here and waited for a couple of hours for refrast to render a frame, so by my reckoning that would still put you at more than 1 minute per frame - hardly what I would call interactive

 

[PC-Engine]

PowerVR Series 5 is expected to be available this year.

 

[Grall]

I've always been a PvR fan (my first 3D card was the venerable Videologic PCX1 daughtercard, which I later fried trying to overclock it lol), but to be quite frank, I'll believe that when I see it.

Been waiting for what feels like decades for a new PowerVR chip and nothing ever seems to happen.

*G*

 

[Kristof]

PowerVR Series 5 is expected to be available this year.

I assume this is based on JohnMs statement in EETimes, lets repeat that statement :

Metcalfe added that the Series-5 architecture would debut in 2003 in a 0.13-micron process technology. He said that it was not yet decided whether Imagination would get first silicon implementations or test chips made at a licensee's wafer fab or at a foundry wafer fab in Taiwan.

So off to dictionary.com to look up "debut" :

debut

n : the act of beginning something new; "they looked forward to the debut of their new product line" [syn: first appearance, launching, unveiling, introduction, entry] v 1: present for the first time to the public; "The and debuts a new song or two each month" 2: appear for the first time in public; "The new ballet that debuts next months at Covent Garden, is already sold out" 3: make one's debut; "This young soprano debuts next months at the Metropolitan Opera"

Presenting for the first time to the public... that could mean anything, from being available all the way down to a paper launch... not saying anything just making sure statements are looked at correctly and nothing is made up or assumed that was not really said 8)

K-

 

[PC-Engine]

Hehe thanks Kristof

It will have PS/VS 3.0 though...

 

[Arun]

Yes, but it won't play games The Way They Are Meant To Be Played, Kristof! j/k

The R420 is likely to be launched ( & available ) after the NV40. So, while I doubt PowerVR will manage to get it ready before nV, if they debut this year, I suppose it'll be at Comdex... So on an avaibility POV, you guys might beat ATI!
And then you've also got to realize ATI will really be using the cheapest ways possible to get Shaders 3.0. compliancy - P Buffer looping to the maximum, slow implementations of branching, ...

So... PowerVR will have a more real VS3.0. / PS3.0. than ATI in H1 2004... Woah...
Now you guys better deliver, or I'm gonna be real decieved, hehe.


Uttar

 

[gkar1]

Yes, but it won't play games The Way They Are Meant To Be Played, Kristof! j/k

The R420 is likely to be launched ( & available ) after the NV40. So, while I doubt PowerVR will manage to get it ready before nV, if they debut this year, I suppose it'll be at Comdex... So on an avaibility POV, you guys might beat ATI!
And then you've also got to realize ATI will really be using the cheapest ways possible to get Shaders 3.0. compliancy - P Buffer looping to the maximum, slow implementations of branching, ...

So... PowerVR will have a more real VS3.0. / PS3.0. than ATI in H1 2004... Woah...
Now you guys better deliver, or I'm gonna be real decieved, hehe.


Uttar

You know what they say when you assume too much

 

[Heathen]

The R420 is likely to be launched ( & available ) after the NV40. So, while I doubt PowerVR will manage to get it ready before nV, if they debut this year, I suppose it'll be at Comdex... So on an avaibility POV, you guys might beat ATI!
And then you've also got to realize ATI will really be using the cheapest ways possible to get Shaders 3.0. compliancy - P Buffer looping to the maximum, slow implementations of branching, ...

Do you have any real proof or are these just complete guesses?

 

[Rodéric]

H1 2004 we'll have PS/VS 3.0 hardware available.

 

[Nick]

I hope it's more than 100 times faster than REF if you want to call it interactive - I've sat here and waited for a couple of hours for refrast to render a frame, so by my reckoning that would still put you at more than 1 minute per frame - hardly what I would call interactive

It's only a guess, but it could be faster. A basic ps 2.0 shader takes about 100 clock cycles per pixel on my renderer. So a complex ps 3.0 shader could be 1000 clock cycles. So a 3 GHz processor could do three million pixels per second. At a reasonable resolution that's three frames per second. And with an overdraw of only three that's one frame per second. I'll call that 'at the limit of being responsive'.

But most of the time developers just want to test one effect, with a simple scene and at low resolution. If the REF needs 10 seconds per frame and I do it at 10 frames per second, then my implementation is a lot more useful isn't it? Also, for CAD purposes a low resolution and framerate can be acceptable, but you still want the fastest implementation...

But actually I do it for the fun of it and I would already be happy with twice the REF performance.

 

[Nick]

H1 2004 we'll have PS/VS 3.0 hardware available.

That should be enough time to get most of it working (the new branch control instructions). The hardest part is still those damn dsx/dsy instructions. Some things need to be redesigned for that... unless someone knows a straightforward way of implementing them and make them work with branch control?

Thanks guys!

 

[DegustatoR]

Shaders 3.0 hardware will be introduced this year.

Street availability is another story. It depends on many factors.

 

[Ailuros]

Shaders 3.0 hardware will be introduced this year.

Street availability is another story. It depends on many factors.

To that I wholeheartedly agree.

Uttar,

Which of them three has taped out yet? That would answer any possible assumption w/o any reliable data in one´s hand and that goes for all three I guess.

Tape outs should be close.

 

[KimB]

Yes, for availability this year, the tapeouts had better be, at the very latest, within a month.

 

[Pavlos]

Nick,
A straightforward (and I think the optimal) solution is to shade a grid of pixels at once. The implementation of partial derivatives (dsx/dsy) becomes trivial but a bit-mask of "active pixels" must be maintained for branching. The opcodes operate or update only the active pixels. Here is a simple and unoptimized implementation of dsx using C:

void dsx(vector** output, vector** input){
	for(int i=0;i<NUM_PIXELS_X;i++)
		for(int j=0;j<NUM_PIXELS_Y;j++)
			if(is_active[i][j] && is_active[i+1][j])
				output[i][j]= input[i+1][j]- input[i][j];
			else
				output[i][j]= 0;	//singularity
}

And here is a reference implementation of the opcode add

void add(vector** output, vector** op1, vector** op2){
for(int i=0;i<NUM_PIXELS_X;i++)
		for(int j=0;j<NUM_PIXELS_Y;j++)
			output[i][j]=(is_active[i][j])?op1+op2: output[i][j];
}

The matrix (or bitmask) is_active[][] must be updated every time the execution path is entering or exiting a conditional expression.

I’m not sure if you can adapt this approach in your framework, but I think you must give it a try. Note that this way has many performance advantages over the other approach (shade each pixel independently).

The SRT Rendering Toolkit

 

[PaulS]

The NV40 was supposed to have taped out already. If it hasn't done so (and we'll possibly get some confirmation on that from nVidia soon, in their conference call on the 7th), then it will have to pretty soon to stand any chance of being available this year.

I don't expect the R420 to tape out soon.

 

[Joe DeFuria]

The NV40 was supposed to have taped out already. If it hasn't done so (and we'll possibly get some confirmation on that from nVidia soon, in their quarterly review), then it will have to pretty soon to stand any chance of being available this year.

Rumors at one point said that the NV40 had taped out...but then the rumors were "corrected", and that it was NV36/38 that had taped out instead.

Unless you have inside info that says otherwise, the current belief is that NV40 had not yet taped out.

I don't expect the R420 to tape out soon.

I expect both the R420 and NV40 to tape out in a similar time frame...within a month or so of each other. I also don't expect any volume shipment of either part to occur until late 1Q '04.