new patent

[Razor1]

new patent?

Embodiments of the present invention provide a method and system for implementing multiple high precision and low precision interpolators in a raster stage of a graphics pipeline. Embodiments of the present invention direct parameter computations requiring high precision to execute on high precision interpolators. Those pixel parameter computations that do not significantly benefit from high precision are directed to execute on low precision interpolators. Both the high precision and low precision interpolator computations execute in parallel. In this manner, embodiments of the present invention provide a graphics processor architecture with reduced power and size requirements.

http://patft.uspto.gov/netacgi/nph-Parser?Sect1=PTO2&Sect2=HITOFF&p=1&u=%2Fnetahtml%2FPTO%2Fsearch-bool.html&r=10&f=G&l=50&co1=AND&d=PTXT&s1=nvidia&OS=nvidia&RS=nvidia

Sorry had the title wrong, should have been a question mark

 

[obobski]

That was filed over 2 years ago, and is just now approved, but wth does it mean, directly, that it's a multi-stage ROP capable of directly writing/reading RAM from different stages?

Also, wouldn't they already use this technology in some form or another?

 

[a688]

Also, wouldn't they already use this technology in some form or another?

Why? A lot of companies come up with ideas that they never end up using or use in the (sometimes far) future. It pays to think ahead.

 

[Geo]

What exactly does DX10 require to be FP32? This patent appears to be a sort of automating of _pp, by class of object/function being gone after. Since it seems unlikely they'd break the spec, that suggests to me it is most likely aiming at those things the spec doesn't require to be FP32. Usually when hearing about DX10, or even DX9, precision, the granularity doesn't get described at a level much below "pixel shader", "vertex shader", or possibly "TMU". I don't see much point to this patent unless there is indeed further granularlity that would allow them to use these lower precision interpolators that the patent is pointed at.

Unless, of course, FP32 is the "lower precision", and they are bringing higher precision for some parts? Would that seem reasonable/useful?

 

[KimB]

Unless, of course, FP32 is the "lower precision", and they are bringing higher precision for some parts? Would that seem reasonable/useful?

Well, from what I understand, the ALU's themselves are only a small fraction of the die area, so it might not be that much of a cost to implement 64-bit floating point. That level of precision would be very useful for GPGPU stuff, and would also be useful for some offline rendering acceleration.

In order for the die area cost to not be too high, one would have to accept that there would be a significant performance hit from increased register pressure and having fewer pixels in-flight when FP64 precision was used.

 

[Arun]

D3D10.1's precision is ">FP32", according to some early specs. G80 is supposedly FP40 but with FP32 registers; it is unknown at this point whether the registers can be FP40 or FP64. So I'd say this patent would still make sense, although I question whether it's even worth for them to bother about...

Uttar

 

[nthd]

I don't see much point to this patent unless there is indeed further granularlity that would allow them to use these lower precision interpolators that the patent is pointed at.

Unless, of course, FP32 is the "lower precision", and they are bringing higher precision for some parts? Would that seem reasonable/useful?

Apologies for the emphasis. There are a great number of patents, quite probably the majority, that in truth have no directly useful purpose. Usefulness is not required for patentability. Companies file vast quantities of applications in the hope that someone, someday, will file something that 'encroaches' on their precious piece of paper. Then, well we all know what happens.

I do not know the details of this patent yet, but I will have a look. Recently I found myself looking at a granted US patent that seemed to be an attempt to claim that a bulletin board delivered over the internet/www was a patentable 'device'. Personally, I much prefer that patents should only be granted on physical applications - not necessarily yet built, but sufficiently well specified to enable such construction - and that copyright laws should hold in the world of programming.

 

[nthd]

It looks to me as if the inventor wants to play 3d games on a mobile phone.

I see no particular discussion of extending precision beyond 32bit - although naturally that is not excluded - but of using fewer bits to achieve superficially similar results to a 'full-sized' operand. The reduction in register size means fewer transistors and hence less power.