FarCry Performance Revisited: ATI Strikes Back with Sh 2b

[radar1200gs]

Thats not the experience I've had with my humble little 5900XT. It performs very nicely with everything I've thrown at it, and far from being forgotten by nVidia in favor of nV40 its performance is still improving (along with IQ) with newer drivers.

 

[Scali]

NV3x supports far more of the spec than R300.

Interesting. What features would that be, for example?
I only know the features that NV3x doesn't support, like multiple render target and floating point textures. But those aren't important I suppose.
NV3x supports multiple element targets... but that's not important either, if multiple rendertargets aren't, I suppose.
What (important) features does the NV3x support that R3x0 doesn't?

 

[Skinner]

I own both X800Xt PE and GF6800U and I'm still amazed how much faster the X800XT PE runs at the indoor shaderscenes/levels (SM3.0 vs SM2.b)

Strange enough its the other way arround in the Source VST with the GF6800U coming out on top in the heaviest part of the demo (with 6600 drives)

 

[Mintmaster]

I was expecting a reply like this...

I suggest you compare the DX9 specifications with the capabilities of NV3x and R3xx. NV3x supports far more of the spec than R300.

I don't know what spec your reading, but NVidia didn't even support FP textures (and not even 16-bit integer textures) until Microsoft supposedly made the special format D3DFMT_A16B16G16Rfc format, which can't do mip-mapping, cubemapping, wrapping, or tiling. I read it in a Meltdown 2003 slide (linked to by someone here at B3D), but can't find any mention of the format even under google. I'm still unsure about whether it exists or not

Without floating point textures, the whole floating point pipeline is almost useless in DirectX. You can't get FP values in via textures, and can't get them out via rendertargets. Therefore no HDR in directx. You get increased precision in a shader, but that's close to meaningless if none of your sources are high precision. Only texture coordinates, and that's it.

R300 didn't come close to NV3x in terms of DX9 support? Get your head out of the sand, buddy. You don't know jack about NV3x's DX9 support.

DX9 without FP barely gives you anything more than DX8 - just longer shaders and a larger instruction set. And when it does, it usually runs like crap on the FX. The only time it doesn't is when NVidia's driver team is able to convert a game' DX9 shaders by hand to DX8-esque shaders, and that'll be done only if the game is popular. As we see in FarCry, they often fail there as well.

The 6000 series is a different story, but the FX is a piece of crap and nothing more.

 

[Ailuros]

Thats not the experience I've had with my humble little 5900XT. It performs very nicely with everything I've thrown at it, and far from being forgotten by nVidia in favor of nV40 its performance is still improving (along with IQ) with newer drivers.

Same goes for competing products of the former generation. The point in case here really is that if the entire NV3x line would have had similar arithmetic efficiency as the NV4x line, yet in analogy to it's specifications, not only would there be much less reason for complaints, but ATI wouldn't had found itself in such a favourable situation last year either.

Some consider it a stupid point, yet the R3xx line had a much higher clock for clock performance/efficiency than the NV3x line had. Tables have obviously reversed nowadays, but it doesn't make it less true for either case or IHV, not in the past and not today.

 

[Apple740]

Strange enough its the other way arround in the Source VST with the GF6800U coming out on top in the heaviest part of the demo (with 6600 drives)

Are you 100% sure that the 6800U walks the DX9.0 path with these 66.00 beta's and not the 'mixed mode' path?

 

[Pete]

According to Valve's own docs, it should. Dave posted a Valve spreadsheet detailing what mode each GPU uses, and the 6800 series uses the DX9 path.

 

[Demirug]

Strange enough its the other way arround in the Source VST with the GF6800U coming out on top in the heaviest part of the demo (with 6600 drives)

Are you 100% sure that the 6800U walks the DX9.0 path with these 66.00 beta's and not the 'mixed mode' path?

There is no "mixed mode" anymore. The 6800 have to render the same shaders as the X800.

 

[pino]

There is no "mixed mode" anymore. The 6800 have to render the same shaders as the X800.

Are you really sure about that ?
I can't believe all HL2 shaders requires more than FP16. It is stupid to run everything with FP32 on NV40 unless you want to slow it down.

Even 3Dmark05 will have PP enabled by default and will use fp16 where is enough.

and then can we really blame NV when they will have to do some shader replacement with drivers ?

 

[Ostsol]

IIRC, mixed mode included both PS2.0 and PS1.x shaders and was primarily meant for the NV3x series -- especially the NV30/31/34, since they still had dedicated integer processing units. AFAIK, partial precision hints are already included in the pure PS2.0 path.

 

[pino]

IIRC, mixed mode included both PS2.0 and PS1.x shaders and was primarily meant for the NV3x series -- especially the NV30/31/34, since they still had dedicated integer processing units. AFAIK, partial precision hints are already included in the pure PS2.0 path.

i see, make sense.

 

[Demirug]

IIRC, mixed mode included both PS2.0 and PS1.x shaders and was primarily meant for the NV3x series -- especially the NV30/31/34, since they still had dedicated integer processing units. AFAIK, partial precision hints are already included in the pure PS2.0 path.

"mixed mode" use PS 1.X and PS 2.X (compiled with the 2.A profile).

The pure PS 2.0 path that is the only DX9 path today do not contain any partial precision hin at all. Tested with one of my tools.

If I add partial precision hints to all shader my GF5900 runs the VST in DX9 mode up to 50% faster. To make sure that the VST renders everything (water) I fake an NV40 device id.

 

[Bjorn]

If I add partial precision hints to all shader my GF5900 runs the VST in DX9 mode up to 50% faster. To make sure that the VST renders everything (water) I fake an NV40 device id.

Interesting. Would be nice with some IQ tests with the DX9, FP 16 mode. And perhaps even more interesting, how does forcing FP16 affect the performance of the NV40 ?

 

[radar1200gs]

It comes as no suprise that Valve considers partial precision to be "non pure 2.0".

I guess it was the least they could do for ATi after the $5 million "donation" and subesquent failure to get the game out the door in a timely fashion...

 

[Bjorn]

"mixed mode" use PS 1.X and PS 2.X (compiled with the 2.A profile).

The "mixed mode" (DX8/DX9) path is the same as the DX8.1 then ?

 

[Unknown Soldier]

I guess it was the least they could do for ATi after the $5 million "donation" and subesquent failure to get the game out the door in a timely fashion...

*Yawn*

Are you really sure about that ?
I can't believe all HL2 shaders requires more than FP16. It is stupid to run everything with FP32 on NV40 unless you want to slow it down.

Even 3Dmark05 will have PP enabled by default and will use fp16 where is enough.

Are you sure? Not according to this Interview with Futuremark. 3DMark2005 will support only SM2.0 and higher. Since FP16 isn't a specification of SM2.0 (ok i'm not sure about that 2.0a path) won't this be a problem with NV3x cards running 3DMark2005??

Isn't SM2.0 supposed to support FP24 and 32 FP only??

US

 

[Bjorn]

Since FP16 isn't a specification of SM2.0 (ok i'm not sure about that 2.0a path) won't this be a problem with NV3x cards running 3DMark2005??

Isn't SM2.0 supposed to support FP24 and 32 FP only??

US

AFAIK, partial precision has been a part of the DX9 SM2.0 specification since it came out. And partial precision is >= FP16. Full precision is FP24 and up though. FP 32 in SM3.0.

 

[pino]

Are you sure? Not according to this Interview with Futuremark. 3DMark2005 will support only SM2.0 and higher. Since FP16 isn't a specification of SM2.0 (ok i'm not sure about that 2.0a path) won't this be a problem with NV3x cards running 3DMark2005??

yeah i am sure, and FP16 is part of sm2.0 :

http://www.beyond3d.com/interviews/fm04/index.php?p=3

3DMark03 obviously caused a lot of contention in some quarters, due to the performance with the PS2.0 tests. Part of the reason for this was that you chose to utilize full precision float shaders and NVIDIA's hardware wasn't optimal in these conditions, despite it being the DirectX default. Would you like to expand on the reasons for this choice? Will 3DMark04 be utilising a mix of full and partial precision shaders in the game tests, dependant on whether the quality allows it? If so, will you be offering options to test partial, mixed and full precision modes such that performances and qualities can be compared?

Patric: Full precision is indeed the DX default and the very few places we used float values in 3DMark03 full precision was needed to keep the image quality. Then again, most materials in ’03 used 1.x shaders that are all fixed point. Now all shaders are SM 2 or 3 and many materials look just the same in full and partial precision. We therefore have a precision mix as default. Materials that get rendered the same in half precision are allowed to use that if the hardware supports it. Then again, if some material gets reduced image quality in half precision, full precision is forced. We plan on adding a switch to force full precision in all shaders for comparison, but that’s only an optional setting.

Isn't SM2.0 supposed to support FP24 and 32 FP only??

No.

 

[Ailuros]

I often wish ATI's next generation part R5xx would be out already. Frankly if there's one thing I can't stand reading anymore are those precision related topics.

You may excuse the interruption.

 

[Demirug]

"mixed mode" use PS 1.X and PS 2.X (compiled with the 2.A profile).

The "mixed mode" (DX8/DX9) path is the same as the DX8.1 then ?

No. The DX8.1 path use PS 1.4 as maximum.

The "mixed mode" was the DX8.2 path.