Uncovering NV35 w/o a review sample...

[Luminescent]

Can anyone with the 5600 ultra benchmark it with v330 of 3DMark03 and report the scores? I am specifically interested in the pixel shader 2.0 test @ 1200*900 resolution.

Supposedly, the 5600 contains 2 fp shading units, 5800 4 fp shading units, and the 5900 8-12 fp units. These extremetech scores depict the fair pixel shader 2.0 results for the 5800 and 5900. I assume that 5600 would scale accordingly with the number of units it has. Being the mark is shader rather than bandwith limited, its performance should reflect well upon the 5800 and 5900's shading performance.

So far we know the 5800 ultra (@500MHz) scores ~11.8 fps while the 5900 ultra (@450MHz) scores ~14.8 fps. The 5600 ultra should then be around fps, assuming 5900 has at least twice the number of units in the 5800.

 

[Hyp-X]

I have 5600 non-Ultra (300MHz) scores at 1024x768 at the moment.
The PS2.0 score is 6.0.

The Ultra is 350MHz so at 1280 it should score 4.48 if it scales linearly (since it's a PS test, that's a good assumption.

That's with 44.03 the drivers which are forcing FP16 on this card.

 

[Luminescent]

Thanks Hyp-X.

If the 5600 ultra scores ~4-5fps @ 1280, we could assume it would be ~6 fps @ 1200*900, which is about half the score of the 5800 ultra. Since both these processors are forcing fp16, the difference between the 5800 ultra and 5900 ultra should less than the difference between 5600 and 5800; 5900 forces fp32.

According to the dawn benchmark results (and thepkrl's NV30 register analysis), the 5900 ultra suffers a ~50% performance gain in going from fp32 to fp16 (18 fps to 27 fps). Taking this into account, the 5900 ultra should (hypothetically) be scoring 14.5*(3/2) or 21.6 fps in fp16 mode with the 3DMark03 pixel shader test. This score is, in fact, close to double the performance of NV30 (@standard clock) and about double (clock for clock). Assuming linear scaling, with a 10% greater clockspeed, NV35's pixel shader score (@500 MHz) should be 24 fps. This score is almost a bit more than double the NV30's score in the pixel shader test.

 

[Luminescent]

I'm beginning to sense we've caught the NV35 with its pants down. It yields almost exactly twice the raw shader performance of NV30, which is capable of 2 fp ops (1 general fmul/add/mad+ 1 fmov) per clock per pipeline.

NV35 was claimed to be capable of 3 fp shader instructions per clock per pipeline or 2 fp shader instructions + 2 independent tex instructions (MDolenc). If it has twice the fp units and it loses an fp unit for every texture op, then it most likely contains 4 fp shaders + 4fp/tex shaders. It has been stated NV35 contains 32 fp units, which divides evenly into 8 sets of 4 fp16/32 clusters. What puzzles me is how NV35 can achieve 3 fp operations per pipeline per clock.

According most NV35 reviews, Nvidia pr states it is capable of ~ twice the shader performance of NV30. The hypothetical results I obtained for 3DMark03, on a completely procedural shader mark, support this.