PDA

View Full Version : FP32 in action in real-world games!


Pages : [1] 2 3 4 5

Hanners
08-13-03, 08:47 AM
I was just reading through Beyond3D's review of the Triplex Radeon 9600 Pro, and came across something that I thought might interest you guys, so here goes:

Beyond3D have now added Tomb Raider: Angel of Darkness to their benchmarking suite, and one of the things they looked at in this review is the difference between the 9600 Pro and the DirectX 9 reference rasteriser. They didn't come up with any differences, except for this one, that shows the difference between FP24 (on the 9600) and FP32 (on the refrast) with the games use of Pixel Shader 2.0 shadows (I've made the screenshots into an animated GIF for comparison):

http://www.hanners.nildram.co.uk/aodfpdemo.gif

The two shots aren't in a 100% identical position, but you can see the added precision of the shadow in the FP32 shot if you look carefully.

Of course this image is zoomed in 7x, so there's no way you'd spot the quality difference in even a normal screen capture, but I thought it was an interesting little example nontheless, more because it's a real-world situation rather than a synthetic benchmark. It's a shame there wasn't a GeForceFX card to throw into the mix too. You can take a look at the part of the review in question here (http://www.beyond3d.com/reviews/triplex/redair9600pro/index.php?p=4).

Nutty
08-13-03, 09:05 AM
That article was strange. It said "Again, there does not appear to be any perceptible differences between the two renderers.", except for the shadow thingy.

1) The shadow in the DX9 reference rasterizer, was a completely different colour to the 9600 image.
2) The Hud health icon, in the DX9 reference shot had an additional glow around the edges, the 9600 image did not.
3) The light glows around the two lights on the Danger sign look completely different.
4) The color of the vent grills was obviously different. The reference render they are light gray, and in the 9600 image almost black. There are numerous instances of this, and not one single reference to it in the text.

They should have done some GF-FX renders also. It would be nice to compare to the reference rasterzier with it.

digitalwanderer
08-13-03, 09:17 AM
It's interesting, but I found TR:AOD to suck so badly that I just don't care how it benches. :)

DaveBaumann
08-13-03, 10:41 AM
Originally posted by Nutty
They should have done some GF-FX renders also. It would be nice to compare to the reference rasterzier with it.

They are very different with these settings.

We may do a comparison article, but we're working through some issues we have with the way the FX boards screw with DOF.

Sazar
08-13-03, 10:47 AM
Originally posted by DaveBaumann
They are very different with these settings.

We may do a comparison article, but we're working through some issues we have with the way the FX boards screw with DOF.

excellent... this sould be an interesting article... show up some expected differences :)

dan2097
08-13-03, 12:07 PM
We may do a comparison article, but we're working through some issues we have with the way the FX boards screw with DOF.

DOF?

Morrow
08-13-03, 12:34 PM
Originally posted by dan2097
DOF?

Depth of Field

IISquintsII
08-13-03, 01:59 PM
Am I blind but all I see is a slight change in the brightness or gamma. Or is this like that story "The Emperors New Clothes" and I am incompetent?

Nutty
08-13-03, 02:18 PM
Probably. The differences I listed are way more obvious, but for some reason the main concern is a shadow that is a few pixels out. For being several thousand times faster than the reference render, I dont particularly care that much, about the shadow being slightly off.

StealthHawk
08-13-03, 05:54 PM
Hmm, but gfFX5200, gfFX5600, and gfFX5800 do not use FP32 in this game, correct? I assume that gfFX5900 does.

serAph
08-13-03, 06:56 PM
whats sad is that fp32 is what the ATi card holders have been dangling above the heads of the 5900 owners for some time now. Good to know it doesnt mean CRAP - unless that animated gif wasnt an adequate demo of the differences.

Hellbinder
08-13-03, 07:24 PM
whats sad is that fp32 is what the ATi card holders have been dangling above the heads of the 5900 owners for some time now. Good to know it doesnt mean CRAP - unless that animated gif wasnt an adequate demo of the differences.

Im really having troubble following your logic here.

1. Nvidia Users are the ones saying that FP32 will be Superior to ATi's FP24.

2. ATi users say "So What, you guys cant run FP32 and will be stuck in FP16 Hell"

3. This test Shows FP32 Rasterizer (Software) V.s 9600pro hardware FP24.

4. It was shown by Nutty that indeed there are Differences between the way FP32 and FP24 look in this case.

5. Extrapolating point 4 = FP16 on Nvidia harware likely to look even Worse.

So please, Tell me. What are you using for the basis of your above statement??

serAph
08-13-03, 07:32 PM
Originally posted by Hellbinder
Im really having troubble following your logic here.

1. Nvidia Users are the ones saying that FP32 will be Superior to ATi's FP24.

2. ATi users say "So What, you guys cant run FP32 and will be stuck in FP16 Hell"

3. This test Shows FP32 Rasterizer (Software) V.s 9600pro hardware FP24.

4. It was shown by Nutty that indeed there are Differences between the way FP32 and FP24 look in this case.

5. Extrapolating point 4 = FP16 on Nvidia harware likely to look even Worse.

So please, Tell me. What are you using for the basis of your above statement??

basis: if the difference between 16 and 24 is as virtually non-existant as 24-32, then bring on the xtra FPS's 16 will give. Or even better, make it an option.

gokickrocks
08-13-03, 09:19 PM
i looked at the full size images of the reference rasterizer and the 9600pro...

am i the only 1 that thinks the 9600 pro is rendering the shadows better? (the shadow im referring to is the shadow of her legs in the full screenshot) [not saying its rendering correctly, but i think its doing a better job]

the reference shot (refer to full size image, not zoomed image) doesnt look like its producing a correct shadow, in fact its brighter than the surroundings (with shadows supposedly being darker than their surroundings)...however on the 9600pro its darker

digitalwanderer
08-13-03, 09:31 PM
Originally posted by serAph
basis: if the difference between 16 and 24 is as virtually non-existant as 24-32, then bring on the xtra FPS's 16 will give. Or even better, make it an option.
There's a much greater visual difference between 16 to 24 as opposed to the slight difference 'tween 24 and 32...it's another reason ATi enthusiasts lord it over the nVidiots. ;)

serAph
08-13-03, 09:48 PM
prove it


- and another thing. ATi fanboys are all like, "the newer games coming out will be run faster by ATi cards." Still havent seen a single instance of this. Honestly, Im open to the proof if its there, but I seriously doubt there IQ difference is noticable and I doubt that the r3xx will ever be significantly better at anything than the FX series - and vice versa.

The Baron
08-13-03, 09:54 PM
Quick!

2^16 = 65536.

2^24 = 16,777,216.

2^32 = 4,294,967,296

So, 24 is a sizable amount. 32, right now, is probably overkill. But 16 is definitely not enough for a lot of things.

(Or so says the man who is spurting out random numbers that don't really have much to do with anything. It's fun, eh?)

Morrow
08-13-03, 10:32 PM
Originally posted by digitalwanderer
There's a much greater visual difference between 16 to 24 as opposed to the slight difference 'tween 24 and 32...it's another reason ATi enthusiasts lord it over the nVidiots. ;)

Is that so because you say it or because it is like that?

I believe it's the former one... ATI fanboys in action trying to save the world which doesn't need a savior.

GeneticWeapon
08-13-03, 10:57 PM
Isnt Reverand working on example of the true differences visually between fp24 and fp32?....maybe it was something he was only pondering....

digitalwanderer
08-13-03, 11:23 PM
Originally posted by GeneticWeapon
Isnt Reverand working on example of the true differences visually between fp24 and fp32?....maybe it was something he was only pondering....
I remember reading him talking about it, but I can't recall.

I read about it in a thread over on Beyond3d that basically was a really long description of the simple math Baron was kind enough to post:

2^16 = 65536.

2^24 = 16,777,216.

2^32 = 4,294,967,296

If you don't understand it, I suggest you go do a search for it on B3D. :)

The long/short is that 65.5K colors isn't enough for right now, 16.7 million is, and 4.2 billion is a tad over-kill-from-hell from where game development is at right now. ;)

The Baron
08-13-03, 11:27 PM
Dig. It's not colors. It's the rounding when using floating point numbers. Big difference ;)

Hellbinder
08-14-03, 12:27 AM
and another thing. ATi fanboys are all like, "the newer games coming out will be run faster by ATi cards." Still havent seen a single instance of this. Honestly, Im open to the proof if its there, but I seriously doubt there IQ difference is noticable and I doubt that the r3xx will ever be significantly better at anything than the FX series - and vice versa.

Why dont you point out a all these "newer games" out there right now that can make a case either way??? Your not seeing the evidence for performance based on "Newer Games" because They Dont Exsist Yet

As it is it was stated by Gabe Newell that HL2's Dx9 shaders run between 5-10x faster on the 9800pro when compared to the 5900U.

Does that mean the game itself will be that much faster? No. But to try to Deny the fact that when running *Standard* Dx9 shaders the Nv30/34/31/35 Get their Asses Kicked is like Denying That the Sky is Blue.

serAph
08-14-03, 12:44 AM
lmfao

you guys crack me up. Lets do some basic arithmetic.

I quote:

2^16 = 65536.

2^24 = 16,777,216.

2^32 = 4,294,967,296

whats 2^24 - 2^16? Regardless of what it is, its nowhere near as big as 2^32 - 2^16, so obviously, its the exact opposite of what your saying, and like Morrow mentioned, the whole lame "nVidia sux, ATi can do no wrong" crap you guys are busy shoveling is really nothing more than what it appears to be: crap.

Like I said, Im still waiting for that awesome reason that the 9800 is soooo much better like you make it sound. Sure the mipping thing sux, but the card is new - give it some time! Im sure that by the time the 5900 is as old as the 9800 is now, the 9800 will be siginificantly slower. Of course, I could be naieve, but then again, history is the ultimate teacher - and look @ the 9800.

DONT bring up that cheating crap either - 45.24 has significant cheating reduction and STILL the 5800 beats the snot out of a 9700. *GASP* Did I just say that?!?! Thats right - and thats In-Game, not friggen 3DMark. The cheats were 3DMark cheats anyway, not all-game specific.

gokickrocks
08-14-03, 01:02 AM
actually all the math you have been doing is wrong in regards to floating point...

fp16 has 1 sign bit, 5 bits for exponent and 10 bits for mantissa
fp24 has 1 sign bit, 7 bits for exponent and 16 bits for mantissa
fp32 has 1 sign bit, 8 bits for exponent and 23 bits for mantissa

Miksu
08-14-03, 01:09 AM
Originally posted by serAph
lmfao


You don't actually read the posts that are written to you or you just love to be a nvidiot. Let me give you a quote:

So, 24 is a sizable amount. 32, right now, is probably overkill. But 16 is definitely not enough for a lot of things.