NVIDIA maps out multi-function interpolators

— 4:12 PM on July 11, 2006

A new paper by a pair of NVIDIA engineers suggests the upcoming G80 graphics processor—or some other "unannounced future product"—may feature "multi-function interpolators." As Elite Bastards explains, interpolation units sit in each pixel shader unit and compute per-pixel attribute values like color, depth, and texture co-ordinates via interpolation. GPUs also contain function evaluators that handle higher-order functions like reciprocal square roots, logarithms, sines, and cosines. However, attribute interpolators and function evaluators are rarely used to their full capacity at the same time, leading to waste of power and die space. To remedy this problem, NVIDIA has apparently merged the two functions in a single unit that can handle both higher-order functions and interpolation. The paper claims this combination improves both performance per die area and performance per Watt.

This sort of sharing of functional units is becoming ever more important in graphics processors. The GeForce 6 and 7-series GPUs have benefited greatly from their use of an ALU in each pixel shader that can also act a texture address unit, raising the GPU's peak potential for both texturing and computation, whichever is in demand. NVIDIA looks to be capitalizing on a similar opportunity for low-level sharing with these new multi-function interpolators. ATI, meanwhile, appears to be focusing on sharing at a higher level, as it has done with the merged vertex and pixel shader units in the Xbox 360 GPU. ATI's upcoming DirectX 10 GPU for the PC is expected to merge vertex and pixel shaders, as well.

Like what we're doing? Pay what you want to support TR and get nifty extra features.
Top contributors
1. BIF - $340 2. Ryu Connor - $250 3. mbutrovich - $250
4. YetAnotherGeek2 - $200 5. End User - $150 6. Captain Ned - $100
7. Anonymous Gerbil - $100 8. Bill Door - $100 9. ericfulmer - $100
10. dkanter - $100
Tip: You can use the A/Z keys to walk threads.
View options

This discussion is now closed.