[volt]

Okay, a friend of mine -- owner of nvision.pl to be exact has just returned from CeBIT. Anyway, he talked to NVIDIA representatives and they told him NV40 will have support for 24-bit floating point precision amongst other things that were confirmed more or less. Below bit is translated (by me) from Polish:

"It's hard to get anything out of NVIDIA representatives. It is obvious they don't want to reveal any information about their next generation GPU. On the other hand, NV40 will premiere very soon and there is already a handful of information available. If you were to ask me about exact specs, I'd have to kill you. However I can always throw in some bits of information. It is worthy to look at current specifications (which were mentioned in our news section before) -- it is very similar to actual NV40 specs. NV40 will also sport hardware MPEG1 / MPEG2 / MPEG4 coding / decoding as well as support for 3D Shadows. NVIDIA has also reworked their 16-bit floating point precision usage -- FP16 and FP32 are said to work flawlessly now. Additionally NV40 will have FP24 support. The above information does look interesting. Did I mention, NVIDIA confirmed 16 pipeline architecture?"

While I talked to Polar about pipeline architecture, he could not confirm full 16 pipeline design. NVIDIA just told him it has 16 pipelines, they did not want to say more. Another important thing is that, NV40 will be widely available after the launch.

Now about R420. He also talked to ATI, but whenever he mentioned the chip, they started to talk about something else. One thing worthy to note is that new drivers will have a totally different look -- a more clean look. Also, ATI did not want to skip driver versions on purpose, someone just "messed it up" -- so no more skipping.

Please note that I did not say any of the above, so don't come after me. It's based on chat I've had with Polar.

[nelg]

Would it not be fair to say if you support Fp32 then by default you also support fP24. After all anything >= Fp24 is full precision.

[Guden Oden]

Why would they bother to support FP24 if FP32 is supposed to be full-speed now? Sounds a bit like those Nvidia guys were BSing on purpose, I have to say.

[Dave Baumann]

IMO there's probably a little confusion thrown in there with full and partial precisions.

[Mintmaster]

I find that odd. If they're going to keep FP32 support, does that really make sense? I mean ATI used FP24 just to save transistors on calculations. They still store textures and framebuffer in FP32 format.

Maybe they converted their FP16 hardware to FP24, and they can output FP16 format like R300 can. Then they can use the faster FP24 whenever they want and the FP32 capability is still there. It still seems a bit silly to me, though.

Or maybe they'll make some demos that instantly show the drawbacks of FP24

[T2k]

I don't understand something: if you have full speed FP32 and now FP24 why you need FP16?

[martrox]

And, If nVidia now has FP24 support...... just what does that do to the FX...... lookks like we could have some serious orphans on our hands......

[991060]

Originally Posted by T2k

I don't understand something: if you have full speed FP32 and now FP24 why you need FP16?

The prolem is that a FP32 chip can get very little performance improvement from using FP24 instead,someone correct me if I'm wrong.

[Mintmaster]

Originally Posted by T2k

I don't understand something: if you have full speed FP32 and now FP24 why you need FP16?

Bandwidth. The same reason R300 can output FP16. Of course, calculations are done internally at another precision.

My guess is this is just a PR mistake, and NV40 does all FP calcs at 32-bit precision internally and writes out whatever format they want.

[991060]

FP16 and FP32 are said to work flawlessly now

Does it mean the register usage problem is gone or alleviated?

as well as support for 3D Shadows

What exactly does it mean? UltraShadow 2?

[Guden Oden]

Originally Posted by Mintmaster

Bandwidth.

Um, whaddya mean bandwidth? R3x0 with its internal 24-bit writes 32-bit pixels externally, so it's not neccessarily a bandwidth saving of any kind...

[digitalwanderer]

"We have now reached infinite improbability!"

[UPO]

My bet is they just thought about full precision (FP32). I wonder about shadow volumes support: would it be possible to calculate silhouette and extrude it on NV40? (faster than on CPU?)

[Tim Murray]

Originally Posted by digitalwanderer

"We have now reached infinite improbability!"

What he said. It won't support FP24--doesn't make any sense considering their PS architecture.

[Ostsol]

Originally Posted by UPO

My bet is they just thought about full precision (FP32). I wonder about shadow volumes support: would it be possible to calculate silhouette and extrude it on NV40? (faster than on CPU?)

Calculating the silhouette kinda has be done in software because not only do the edges have to be found, but new vertices have to be created to define the volume. Extrusion has been done via vertex shaders for quite a while, though.

[mreman4k]

I wonder why you would support FP24 is you can run FP32/FP16 flawlessly. I could have swore that one of NVidia's chief scientist state that FP24 was blasphemy.

Good little tidbit about a cleaner look to ATI drivers

[Tim Murray]

Originally Posted by 991060

Does it mean the register usage problem is gone or alleviated?

I guess I forgot that rumor. Yes, it has a hell of a lot more registers. I've heard one number for the amount of registers, but I haven't gotten any sort of confirmation on it. So, I'll leave it with yes, the register problem is at least alleviated and very possibly gone completely.

[digitalwanderer]

Originally Posted by mreman4k

I wonder why you would support FP24 is you can run FP32/FP16 flawlessly. I could have swore that one of NVidia's chief scientist state that FP24 was blasphemy.

I just can't see them supporting it after all the FUD they pumped out against it over the last year and a half with their "cinematographantasmaorgasmic" talk that they'd just be opening themselves up for some HUGE attacks on their hypocrisy and screwing over the entire installed FX base.

Good little tidbit about a cleaner look to ATI drivers

No, it was an evil tidbit about a cleaner loot to ATi's drivers....it's going to keep me up at night wondering about it!

[Tim Murray]

Look, it doesn't support FP24. I've gotten a good amount of specific PS info, and it has no support for FP24. FP32/FP16.

[digitalwanderer]

Originally Posted by The Baron

Look, it doesn't support FP24. I've gotten a good amount of specific PS info, and it has no support for FP24. FP32/FP16.

And let me guess, you're at school so me turning on Trillian would be a waste of my time.

[Tim Murray]

Originally Posted by digitalwanderer

And let me guess, you're at school so me turning on Trillian would be a waste of my time.

I'll be home in fifteen minutes, Snookums.

[digitalwanderer]

Originally Posted by The Baron

I'll be home in fifteen minutes, Snookums.

Just in time for me to have to go pick up my son from school...kind of weird/scary, eh?

BTW-I checked out nVnews today and almost fell over laughing, there's a thread where some of the most raving nVidia fanboys seem to have found a rumor they believe telling 'em that the nV40 will be 8x2 and the R420 will be a true 16x1 and they're panicking and moaning and already giving the crown to ATi! I almost peed myself laughing, I'll hunt up a link.

EDITED BITS: Found it! "R423 will have 16 pipes(16x1) and will own nv40".

Sorry for the OT, I blame it on a lack-o-new-rumors.

[UPO]

Originally Posted by Ostsol

Calculating the silhouette kinda has be done in software because not only do the edges have to be found, but new vertices have to be created to define the volume. Extrusion has been done via vertex shaders for quite a while, though.

Well actually you don't have to create new vertices on fly. Creating shadow volumes entirely on GPU is possible. The only question is cost.
http://www.mpi-sb.mpg.de/units/ag4/g...stefan.brabec/
http://www.gamedev.net/columns/hardc...lume/page5.asp

[Ostsol]

Originally Posted by UPO

Well actually you don't have to create new vertices on fly. Creating shadow volumes entirely on GPU is possible. The only question is cost.
http://www.mpi-sb.mpg.de/units/ag4/g...stefan.brabec/
http://www.gamedev.net/columns/hardc...lume/page5.asp

Ah! That's true, though in this case speed would be entirely related to vertex processing capabilities, so it's kinda off topic.

[3dcgi]

Originally Posted by T2k

I don't understand something: if you have full speed FP32 and now FP24 why you need FP16?

One reason for FP16 is filtering and blending in FP16 is cheaper than FP32.