Doom 3 - NVidia colour precision - ARB2 vs NV30

[g__day]

JC has stated that on the ARB2 path NVidia runs Doom 3 at about half the speed of the NV30 path. Does this infer that ARB2 path being the OPEN standard is enforcing higher colour precision and NV3x can't optimise its pipelines well (it is stuck in f32 mode killing performance) whilst in NV30 path NV3x can mix and balance fx12, fp16 and fp32 colour precisions to double the number of instructions executed through its rendering pipelines?

 

[Doomtrooper]

Yes

 

[tEd]

There was even a comment on toms which said that the arb2 path will have the same speed as the nv30 path with the final drivers. Coincentally the arb2 path wasn't working right when they did the all famous doom3 benchmarks.

That sentence certainly put a smile on my face

 

[Luke Philpot]

That sentence certainly put a smile on my face

Same here

 

[chavvdarrr]

AFAIK, NV3x receives severe performance hit when increasing number of internal registers are used - looks like the greater speed in fp16 mode is due to the smaller register usage...

 

[g__day]

Isn't this a rather strange limitation - would it just have costed too many transistors from the budget to provide enough internal registers to allow all those pipelines to fly?

Arjans about mid way down this post seemed to think it would take alot or registers at every stage of every pipeline. If so thats alot of overhead.

http://www.beyond3d.com/forum/viewtopic.php?t=6872&start=20

 

[Dio]

Isn't this a rather strange limitation - would it just have costed too many transistors from the budget to provide enough internal registers to allow all those pipelines to fly?

There are many possible architectures... the answer is 'it certainly could have'...