Pipe Confusion
- Thread starter Sinistar
- Start date
Sinistar said:
"Peak fragment calculation performance is 64 fragments per clock cycle"
"Peak pixel output is 8 pixels per clock". In my opinion one have to be more specific... Allthought, for example Nvidia PR managed to make 4 pipe GPU to 8 pipe one in past, so everything is possible
Sinistar said:
If you haven't read any of the reviews, it might be I suppose. You think that likely for purchasers after a $600 card? I suppose it might happen, but not too often. It's not like they are trying to hide it.
They also say "The GeForce 7800 GTX has a 24 pipe architecture compared to a 16 pipe architecture of the 6800 Ultra" in their faq, for instance.
I'd call it "putting their best foot forward", but we should all realize (as Wavey tried to pound home in the C1 article) we are in an era where our old paradigms and language are going to have to evolve in order to communicate effectively.
Pixels output per clock really isn't a good indication of performance, not any more. Much more indicative is the number of pixels worked on in parallel per clock, so while not technically correct, the statement that the 7800 has "Pixels per clock (peak): 24" is not misleading. It's a good way to translate the performance of the part as measured by previous parts that have had ROP's tied directly to pixel pipelines.
Martin Eddy
Regular
NV30/35/38 had 4 pixel pipelines not 8.
NV31/34/36 had 2 pixel pipelines not 4.
NV31/34/36 had 2 pixel pipelines not 4.
As per the currently most accepted definition of pixel, the NV30/NV35/NV38 have a single quad pipeline, while the NV31/34/36 have a single half-quad pipeline with a bypass path.
The number of ROPs is also incorrect, I think it was 8 for the NV30 and 4 for the NV31, but it's been a long time since I checked that
Uttar
The number of ROPs is also incorrect, I think it was 8 for the NV30 and 4 for the NV31, but it's been a long time since I checked that
Uttar
4th different opinion during 24 hours. Nice My first personal opinion was that NV30/35/38 (5800/5900/5950) are 4 pipes (2 TMUs) / 4 ROPs (I agree with you), but NV31/34/36 (5200/5500/5600/5700) use 2 pipes or 2 ROPs (or both) architecture, because they can output only 2 pixels/clock, hmm?
no-X said:
My first personal opinion was that NV30/35/38 (5800/5900/5950) are 4 pipes (2 TMUs) / 4 ROPs (I agree with you), but NV31/34/36 (5200/5500/5600/5700) use 2 pipes or 2 ROPs (or both) architecture, because they can output only 2 pixels/clock, hmm?
Basically, NV31/34/36 pixel shading pipes can work in 2 different ways
4 pixel pipes, 1 TMU, only 1 pipeline pass
or
2 pixel pipes, 2 TMUs, many loopbacks
As far as I know, NV30 and NV35 has:
4 pixel pipelines with 2 texturing units on each, basically 4x2 configuration,
in other words a "1 Quad" pipeline. 4 ROPs, 4x (Color + Z), 8x Z values / clock with no-AA. ( giving it an edge in Doom3 ), 3 vertex units, 8 pixel shader units ( some functional units handling fixed-point math replaced with floating-point units in the NV35 )
Also:
NV31 is 2 pipes with 2 texturing units each, Half-Quad Pipe, 2 ROPs, 2x (Color + Z), 4x Z values / clock with no-AA. ( doing "better" in Doom3 than expected ), derived from NV30, 1 Vertex Unit, 4 pixel shader units.
NV36 is 2 pipes with 2 texturing units each, Half-Quad Pipe, 2 ROPs, 2x (Color + Z), 4x Z values / clock with no-AA. ( doing "better" in Doom3 than expected ) - 3 vertex units, 4 pixel shader units. it`s performing way better than NV31 due to 3x Vertex Shader units and being derived from NV35 it means some functional units that handled fixed-point operations were replaced with floating-point units. ( helping Shader Model 2.0+ shaders, and negatively affecting Shaders 1.1 and 1.4 - where fixed point is used )
NV34 is 2 pipes with 2 texturing units each, Half-Quad Pipe, 2 ROPs, 2x (Color + Z), 2x Z values / clock with or without AA. Derived from NV30, 1 vertex unit, 4 pixel shader units.
6-Series:
NV43 has 8x1 configuration, 2x Quad Pipeline, 1 texturing unit on each, 8 shader units, linked with a crossbar to 4 ROPs, 4x ( Color + Z ), 8x Z operations, 3 vertex shaders : doing very good in Doom3 with noAA, because the 8x Z ROPs optimization. ( NV realized that multitexturing performance has less importance in the Shader Model 2.0+ era - so 8 shaded pixels at the same time gives the edge - also linking 8 shaders to 4 ROPs is not a problem, because usually more than 1 pass is required for a pixel )
NV44 has 4x1 configuration, 1x Quad Pipeline, 1 texturing unit on each, 4 shader units, linked with a crossbar to 4 ROPs, 4x ( Color + Z ), 4x Z operations, 3 vertex shaders.
NV40 has 16x1 or 12x1 configuration, 16 ROPs, 16x ( Color + Z), 32x Z.
7-Series:
NV47 8)
24x1 configuration, 16 ROPs, 16x ( Color + Z ), 32x Z
AFAIK
4 pixel pipelines with 2 texturing units on each, basically 4x2 configuration,
in other words a "1 Quad" pipeline. 4 ROPs, 4x (Color + Z), 8x Z values / clock with no-AA. ( giving it an edge in Doom3 ), 3 vertex units, 8 pixel shader units ( some functional units handling fixed-point math replaced with floating-point units in the NV35 )
Also:
NV31 is 2 pipes with 2 texturing units each, Half-Quad Pipe, 2 ROPs, 2x (Color + Z), 4x Z values / clock with no-AA. ( doing "better" in Doom3 than expected ), derived from NV30, 1 Vertex Unit, 4 pixel shader units.
NV36 is 2 pipes with 2 texturing units each, Half-Quad Pipe, 2 ROPs, 2x (Color + Z), 4x Z values / clock with no-AA. ( doing "better" in Doom3 than expected ) - 3 vertex units, 4 pixel shader units. it`s performing way better than NV31 due to 3x Vertex Shader units and being derived from NV35 it means some functional units that handled fixed-point operations were replaced with floating-point units. ( helping Shader Model 2.0+ shaders, and negatively affecting Shaders 1.1 and 1.4 - where fixed point is used )
NV34 is 2 pipes with 2 texturing units each, Half-Quad Pipe, 2 ROPs, 2x (Color + Z), 2x Z values / clock with or without AA. Derived from NV30, 1 vertex unit, 4 pixel shader units.
6-Series:
NV43 has 8x1 configuration, 2x Quad Pipeline, 1 texturing unit on each, 8 shader units, linked with a crossbar to 4 ROPs, 4x ( Color + Z ), 8x Z operations, 3 vertex shaders : doing very good in Doom3 with noAA, because the 8x Z ROPs optimization. ( NV realized that multitexturing performance has less importance in the Shader Model 2.0+ era - so 8 shaded pixels at the same time gives the edge - also linking 8 shaders to 4 ROPs is not a problem, because usually more than 1 pass is required for a pixel )
NV44 has 4x1 configuration, 1x Quad Pipeline, 1 texturing unit on each, 4 shader units, linked with a crossbar to 4 ROPs, 4x ( Color + Z ), 4x Z operations, 3 vertex shaders.
NV40 has 16x1 or 12x1 configuration, 16 ROPs, 16x ( Color + Z), 32x Z.
7-Series:
NV47 8)
24x1 configuration, 16 ROPs, 16x ( Color + Z ), 32x Z
AFAIK
Alstrong said:Can people just give corrections if they know it's wrong?
It has already been discussed many times. A quick summary :
NV30 :
3 VS, 4 pixel pipes, 4 ROPs
Double Z
pixel pipe = 1 FP32 ALU handling 2 TMUs + 2 FX12 Mini-ALU (each one can do 2 MULs or 1 ADD or 1 MAD)
NV35/38 :
3VS, 4 pixel pipes, 4 ROPs
Double Z
pixel pipe = 1 FP32 ALU handling 2 TMUs + 2 FP32 mini ALU (each one can do 1 MUL or 1 ADD or 1 FP16 MAD)
NV31/34 :
1VS (actually 3 VS for NV31 but 2 were broken), 4 or 2 pixel pipes, 4 ROPs
pixel pipe mode 1 (4 pipes) = 1 TMU (+ 1 mini-ALU but I don't remember all the details)
pixel pipe mode 2 (2 pipes) = 1 FP32 ALU handling 2 TMUs + 2 FX12 Mini-ALU (each one can do 2 MULs or 1 ADD or 1 MAD)
NV36:
3VS, 4 or 2 pixel pipes, 4 ROPs
pixel pipe mode 1 (4 pipes) = 1 TMU (+ 1 mini-ALU but I don't remember all the details)
pixel pipe mode 2 (2 pipes) = 1 FP32 ALU handling 2 TMUs + 2 FP32 mini ALU (each one can do 1 MUL or 1 ADD or 1 FP16 MAD)
