15-Way Open-Source Intel/AMD/NVIDIA GPU Comparison

[curaga]

Just curious, where do you get the ROP specs from? I can't seem to find them even for any Radeon, let alone for Intel chips.

[bridgman]

For AMD parts it's pretty easy -- we release detailed block diagrams to the review sites with each new HW generation and that info ends up on the Wikipedia pages with pretty decent accuracy.

http://www.tomshardware.com/reviews/radeon-hd-6450-caicos-blu-ray-3d,2920.html

For the Intel parts there's a lot more reading between the lines required -- seems to be consensus that the HD 4600 has 2 ROPs but no indication how wide they are. The EUs seem to be 4-wide ALU sets and various slides talk about "hashes" or "quads" of pixels, which argues for 4 pixel wide ROPs (8 total). The raw fill performance seems to support that as well.

To add to the complication, the pixel processing blocks (we call them DB/CB for depth/colour buffer) often include more units for Z & stencil ops than for colour operations, the idea being that a lot of "potential pixels" don't get drawn because they are obscured by already-drawn content (reflected in the Z buffer). The AMD parts typically support 4x the number of Z ops as colour ops, not sure about Intel. The implication of this is that you have to be careful which synthetic benchmarks you look at.

[curaga]

For radeons I meant the ROP width. The number of ROPs is everywhere yes, but nowhere does it say how many pixels each processes per clock.

[bridgman]

For AMD parts the convention seems to be counting ROP width in pixels, ie 4 ROPs = 4 pixels per clock of colour processing (16 per clock for Z or stencil).

There's no formal standard though (and we don't even call them ROPs ).