3.0 PS...kind of a let down in FarCry

[Pete]

NV40 only offers FP16 blending, AFAIK. It must have been a misprint. FP32 blending is not the same as just FP32 precision/render targets. Read any 6800U preview and it'll specify FP16 blending only.

 

[Shad0hawK]

Originally posted by: Pete
NV40 only offers FP16 blending, AFAIK. It must have been a misprint. FP32 blending is not the same as just FP32 precision/render targets. Read any 6800U preview and it'll specify FP16 blending only.

no it is not a "misprint" you got some bad info somewhere. the only card confirmed not to do fp32 is ATI even the fx series can do fp32

it was speciifically mentioned as well at the launch event by cervat yerli, tim sweeny plus even anand even talks about fp32

"Actually, this time around, NVIDIA is supporting front to back fp16 all the way from the software to the framebuffer. This will assist in things like HDR rendering, as the fp16 (or fp32) data calculated in the pixel shaders no longer needs to be converted to 8bit integer color for display. "

here is the link

do you have any credible links stating that the 6800 series does not do fp32?

 

[CaiNaM]

actually, pete is correct here.. blending and rendering are different... as previously pointed out in this thread:

"Now, GeForce 6800 Ultra introduces an optional 64 bit floating point framebuffer. That is, RGBA FP16 FP16 FP16 FP16. We already covered its higher range compared to FX8. While nVidia calls it "high dynamic range" (HDR), it is in fact medium dynamic range (MDR). HDR needs at least 32 bit for each single value, NV40 has 16 bit. Anyway, FP16 is a great leap forward. The main advantage is: no longer do we have to render to an FP texture for MDR rendering. Just activate the FP16 framebuffer and you'll get an aceptable MDR render target. NV40 also supports all the alphablending stuff with its 64 bit framebuffer"

~ 3dcenter

at any rate, it's still a better method than what's avail. on r420 hardware.

 

[BFG10K]

Yes according to the technical definition of HDR nVidia can't do it because it requires FP32 blending. They can do MDR though. To do true HDR you need a FP32 pipeline/frame buffer.

R3xx/R4xx & NV3x cards can probably do "HDR" too but it'll be slower and hacked around a bit compared to the NV4x.

 

[Shad0hawK]

Originally posted by: CaiNaM
actually, pete is correct here.. blending and rendering are different... as previously pointed out in this thread:

"Now, GeForce 6800 Ultra introduces an optional 64 bit floating point framebuffer. That is, RGBA FP16 FP16 FP16 FP16. We already covered its higher range compared to FX8. While nVidia calls it "high dynamic range" (HDR), it is in fact medium dynamic range (MDR). HDR needs at least 32 bit for each single value, NV40 has 16 bit. Anyway, FP16 is a great leap forward. The main advantage is: no longer do we have to render to an FP texture for MDR rendering. Just activate the FP16 framebuffer and you'll get an aceptable MDR render target. NV40 also supports all the alphablending stuff with its 64 bit framebuffer"

~ 3dcenter

at any rate, it's still a better method than what's avail. on r420 hardware.

any links other than to other threads on forums?

 

[Shad0hawK]

Originally posted by: Pete
NV40 only offers FP16 blending, AFAIK. It must have been a misprint. FP32 blending is not the same as just FP32 precision/render targets. Read any 6800U preview and it'll specify FP16 blending only.

ok i am with you now

here is a quote from the Neoseeker interview with David Kirk

So, for real addressing or geometric math, 24bit FP is not enough. On the flip side, for color calculation, 24bit FP is overkill. 16bit FP is precise enough to represent color/brightness range of approximately 14dB, which is almost the full range of the human visual system. So, 16bit FP is also a convenient format for storing and filtering textures and blending images. As further support for that idea, ILM and other major film studios have been using it for years.

*edit*

so basically what nvidia has done is put a full fp32 pipeline, but used fp16 to do the color calculation because that is basically the range of human vision anyway, but the room to expand is there regardless. so while the color calculation is done if fp16 it can be rendered/processed at fp32 without the performance hit of losing half the available (pipeline)bandwidth doing the color calculations in fp32.

cool

 

[Shad0hawK]

Originally posted by: BFG10K
Yes according to the technical definition of HDR nVidia can't do it because it requires FP32 blending. They can do MDR though. To do true HDR you need a FP32 pipeline/frame buffer.

R3xx/R4xx & NV3x cards can probably do "HDR" too but it'll be slower and hacked around a bit compared to the NV4x.

the 6800 does have a full fp32 pipeline.(which was one of the reasons the die was so big)

 

[CaiNaM]

Originally posted by: Shad0hawK

Originally posted by: BFG10K
the 6800 does have a full fp32 pipeline.(which was one of the reasons the die was so big)

well, that's not really correct... at least in the context of this discussion. yes, each pixel shader 'pipeline' (the most commonly misused/misunderstood term regarding these cards) does have an fp32 texture unit, however the fp32 (128 bits) refers to math ops and the fp16 we're talking about here refers to color values (texture & frame buffer blending) as it applies to dynamic range lighting.

bottom line is it CAN be done on either nv40 or r420, but theoretically it will be done easier/cleaner and at less the performance cost with the nv40... tho we really need to see something like hl2 (will it EVER ship?) which supports H*cough*DR lighting on both r420 and nv40 to compare any performance or quality differences.

 

[Safeway]

This is ridiculous. 3.0 was never going to increase eye candy. 2.0 can do everything 3.0, just a little slower.

 

[Keysplayr]

Originally posted by: Safeway
This is ridiculous. 3.0 was never going to increase eye candy. 2.0 can do everything 3.0, just a little slower.

Actually, if you go up and take a look at the link I posted for the review, some extra items did show up in one of the screenshots from the NV40 using the latest nv drivers and the 1.2 SM3.0 patch in FarCry.
Many visual anomalies were corrected if not all on the nv40 so it is now on par IQ wise with ATI. Lighting saturation is less pronounced in SM3.0 which may not mean anything imparticular. Other than that one screenie where extra items showed themselves, I didn't see anything extra either. 3.0 is just to increase performance by decreasing the time it takes to render. Do you agree?

 

[Shad0hawK]

Originally posted by: Safeway
This is ridiculous. 3.0 was never going to increase eye candy. 2.0 can do everything 3.0, just a little slower.

actually that depends on the number of effects used, according to Cervat Yerli in some cases they had hit the maximum numbers of instructions SM2 could do. sm3 can do a tremendous amount more, which is why the used sm3 in the first place.

or to put it in perspective, a pentium4ee 3.4 ghz does the same thing a pentium4 2.8 ghz does....

 

[CaiNaM]

in all fairness, keep in mind tho that ati has exteneded the insturctions limit of sm2 on r420. it's unlikely the limit will be reached on r420 in the near future.

if nv40 does more effects, it will be more to due each effect having less impact on performance rather than any limitation in the # of instructions.

 

[Shad0hawK]

Originally posted by: CaiNaM

Originally posted by: Shad0hawK

Originally posted by: BFG10K
the 6800 does have a full fp32 pipeline.(which was one of the reasons the die was so big)

well, that's not really correct... at least in the context of this discussion. yes, each pixel shader 'pipeline' (the most commonly misused/misunderstood term regarding these cards) does have an fp32 texture unit, however the fp32 (128 bits) refers to math ops and the fp16 we're talking about here refers to color values (texture & frame buffer blending) as it applies to dynamic range lighting.

bottom line is it CAN be done on either nv40 or r420, but theoretically it will be done easier/cleaner and at less the performance cost with the nv40... tho we really need to see something like hl2 (will it EVER ship?) which supports H*cough*DR lighting on both r420 and nv40 to compare any performance or quality differences.

well until isee any evidence other than forum posts i will believe this:

"Nvidia stuck by their guns however and have a full fp32 pipeline implemented through the GeForce 6800 Ultra. According to Jen-Hsun implementatin of fp32 over fp24 costs Nvidia a whopping 25% more in terms of die space but they feel that they will be vindicated with the benefits down the road and do not regret the decision to go with fp32 with the NV3x series. "

linky

and

"The FP32 texture unit found in each pixel shader pipeline also tends to filtering chores, doing bilinear, trilinear, and up to 128-tap anisotropic filtering. However, these texturing units can also perform filtering on FP16 color values, such as those used in ILM's OpenEXR format. This preserves pixel color precision by not "dumbing down" the pixel color value to a fixed point 32-bit value, where filtering operations could introduce rounding errors. These errors can show up as banding or blotching in the image, particularly in areas with higher than normal dynamic range. It also means that an FP16 color value can be written into the frame buffer, and then be read back into the GPU without any loss of precision."

and combined with my earlier comment...about fp32 processing/fp16color bending i do not see anything factually incorrect,

Originally posted by: Shad0hawKso basically what nvidia has done is put a full fp32 pipeline, but used fp16 to do the color calculation because that is basically the range of human vision anyway, but the room to expand is there regardless. so while the color calculation is done if fp16 it can be rendered/processed at fp32 without the performance hit of losing half the available (pipeline)bandwidth doing the color calculations in fp32.

 

[Safeway]

Yea, I understand that, which is why I said 2.0 can do everything 3.0 can (and by that, I mean presently, since current shader code is about a 10th of 2.0 max), only 2.0 is slower. I think people were expect way way way improved IQ, which was caused by not understand what the 3.0 was bringing. And by eye candy, I meant IQ not related to FPS. AA/AF isn't improved, which I actually read posted on another forum that it would be. Guess what, 3.0 cures genital herpes too! ...

It isn't a god send, it's a nice feature that will be useful once native 3.0 games come out (soon). Also, I am game makers will continue to write 2.0, it will just run mildly slower than 3.0.

 

[CaiNaM]

Originally posted by: Shad0hawK

Originally posted by: CaiNaM

Originally posted by: Shad0hawK

Originally posted by: BFG10K
the 6800 does have a full fp32 pipeline.(which was one of the reasons the die was so big)

well, that's not really correct... at least in the context of this discussion. yes, each pixel shader 'pipeline' (the most commonly misused/misunderstood term regarding these cards) does have an fp32 texture unit, however the fp32 (128 bits) refers to math ops and the fp16 we're talking about here refers to color values (texture & frame buffer blending) as it applies to dynamic range lighting.

bottom line is it CAN be done on either nv40 or r420, but theoretically it will be done easier/cleaner and at less the performance cost with the nv40... tho we really need to see something like hl2 (will it EVER ship?) which supports H*cough*DR lighting on both r420 and nv40 to compare any performance or quality differences.

well until isee any evidence other than forum posts i will believe this:

"Nvidia stuck by their guns however and have a full fp32 pipeline implemented through the GeForce 6800 Ultra. According to Jen-Hsun implementatin of fp32 over fp24 costs Nvidia a whopping 25% more in terms of die space but they feel that they will be vindicated with the benefits down the road and do not regret the decision to go with fp32 with the NV3x series. "

linky

and

"The FP32 texture unit found in each pixel shader pipeline also tends to filtering chores, doing bilinear, trilinear, and up to 128-tap anisotropic filtering. However, these texturing units can also perform filtering on FP16 color values, such as those used in ILM's OpenEXR format. This preserves pixel color precision by not "dumbing down" the pixel color value to a fixed point 32-bit value, where filtering operations could introduce rounding errors. These errors can show up as banding or blotching in the image, particularly in areas with higher than normal dynamic range. It also means that an FP16 color value can be written into the frame buffer, and then be read back into the GPU without any loss of precision."

and combined with my earlier comment...about fp32 processing/fp16color bending i do not see anything factually incorrect,

Originally posted by: Shad0hawKso basically what nvidia has done is put a full fp32 pipeline, but used fp16 to do the color calculation because that is basically the range of human vision anyway, but the room to expand is there regardless. so while the color calculation is done if fp16 it can be rendered/processed at fp32 without the performance hit of losing half the available (pipeline)bandwidth doing the color calculations in fp32.

well, first, you're taking my comments out of context. i quoted an entirely different statment. but if you want to split hairs, the above statement is not entirely accurate either, as there is no "room to expand". nv40 is not capable of higher than fp16 fb blending (tho fp16 blending IS a good thing). you keep combining color calc with math ops, when they are seperate entities...while they are partof the same overall equation, they are entirely different terms.

 

[Shad0hawK]

Originally posted by: CaiNaM
well, first, you're taking my comments out of context. i quoted an entirely different statment. but if you want to split hairs, the above statement is not entirely accurate either, as there is no "room to expand". nv40 is not capable of higher than fp16 fb blending (tho fp16 blending IS a good thing). you keep combining color calc with math ops, when they are seperate entities...while they are partof the same overall equation, they are entirely different terms.

so how am i "taking them out of context?" i even differentitated between the two (the exact opposite of combining i believe)before you came in and made this the long winded afair it is now?

all i am asking is for some links so i can do some more reading on it, some people on forums seem to be saying the nv40 does not have a fp32 pipeline, tech articles i have read say it does...i know which carries more weight with me.

 

[BFG10K]

the 6800 does have a full fp32 pipeline.(which was one of the reasons the die was so big)

You are confusing the math ops (FP32) with colour/blending ones (FP16).

 

[Shad0hawK]

Originally posted by: BFG10K

the 6800 does have a full fp32 pipeline.(which was one of the reasons the die was so big)

You are confusing the math ops (FP32) with colour/blending ones (FP16).

actually, i am not. i am very well aware that the colors are done in fp16 since last night(i even elucidated in a post). while the article i quoted mentioned that fp16 color blending is near the human limit of discernment so anything above that is primarily a waste anyway. it also mentioned ILM uses fp16 color blending in the work they do as well, which i though was very cool.

i would be very happy to believe you, but the tech sites i see and nvidia's chief engineer call it a "full fp32 pipeline" if you guys have any links stating otherwise(not to to other forums or posts that is) i would be happy to read them.

 

[CaiNaM]

Originally posted by: Shad0hawK

Originally posted by: BFG10K

the 6800 does have a full fp32 pipeline.(which was one of the reasons the die was so big)

You are confusing the math ops (FP32) with colour/blending ones (FP16).

actually, i am not. i am very well aware that the colors are done in fp16 since last night(i even elucidated in a post). while the article i quoted mentioned that fp16 color blending is near the human limit of discernment so anything above that is primarily a waste anyway. it also mentioned ILM uses fp16 color blending in the work they do as well, which i though was very cool.

i would be very happy to believe you, but the tech sites i see and nvidia's chief engineer call it a "full fp32 pipeline" if you guys have any links stating otherwise(not to to other forums or posts that is) i would be happy to read them.

lol.. you're completely missing the point... what started this ridiculous point you're trying to argue is"

bfg: Yes according to the technical definition of HDR nVidia can't do it because it requires FP32 blending. They can do MDR though. To do true HDR you need a FP32 pipeline/frame buffer.

tho which you quoted him and replied: the 6800 does have a full fp32 pipeline.(which was one of the reasons the die was so big)

to which we are both saying your statement saying, "6800 does have a full fp32 pipeline" has nothing to do with it's inability to do HDR.. it has to do with the framebuffer, not the pipeline....

not sure about bfg, but i'm sure trying to figure out what you're arguing. neither of us ever said nv40 didn't have a fp32. in fact one of my specific statements was, "each pixel shader 'pipeline' (the most commonly misused/misunderstood term regarding these cards) does have an fp32 texture unit"...

you keep quoting out of context, applying one reply to a compeltely different quote, then arguing something which has nothing to do with the other... the correct information was given, and there's not further point in engaging in a discussion you don't seem to want to keep in context.