Printable View
I guess Nvidia have a pile of parts that were not up to the grade of being a GTX 470 or 480.
http://www.fudzilla.com/content/view/18390/1/
This ties in with the "Custom GTX 480 boards" that seem to be popping up.
http://www.brightsideofnews.com/news...elopments.aspx
You think those "custom 480" boards are really for the 460? Makes way more sense to me.
Yeah I think that makes more sense with respect to the 6+6 pin arrangement certainly, though I'd be surprised if nVidia were only letting them develop PCBs for one variant when the layout is the same.
i hope the gtx 260 performs well, and i am not sure how the custom gtx 480 would work...cramping all these stuff in a small pcb would make it even hotter than what it is now
Makes perfect sense if you assume the custom boards aren't actually for making a 480.Quote:
Originally Posted by vian_siu
Nvidia must be sitting on a mountain of silicon that didn't make the grade to become a 470 let alone a 480.
If the 460 has less streams in use it can run cooler than a 480, but uses the same chip as a 480, so a smaller board is more profitable for no downside.
I wonder how badly they'd have to gimp the chip to get it working cool enough / at a low enough power draw - the rumours were that the GTX480 runs so hot because the defects in manufacturing are spread evenly through the silicon which mean that even if you fuse off the worst of the defects you still need to notch the voltage up to run stably.
I suppose if they could find some silicon where they could fuse off 4 shader clusters they might have a marketable card @ 384 shaders at around the same clocks as a GTX470. The problem there is that it would have 15% less shaders (and hence performance) than a GTX470, and the 470 doesn't have that kind of consistent lead over the 5850. It has it in some benchmarks, but in others it's barely ahead at all. Also, nvidia's usual approach to lower end cards is to cut both shaders and memory interface, and I could envisage a GTX460 being cut back to 384 shaders *and* 1GB DDR5 on a 256bit interface. Given that nvidia don't seem to be able to run DDR5 stably above about 900MHz, that'd really cut into the available memory bandwidth...
I got the impression they had to up the voltage enough to not only get the chip basically stable but then up another notch in order to increase the clock speed to compensate for the fused off shaders.
They must be faced with a growing pile of silicon that failed validation to go into a GTX470, and I expect they are playing with options as we speak as to how they best re-validate that silicon for a new part. Do they fuse off lots of clusters to get good yield, or a few which means they won't salvage as many parts? If they fuse off lots of clusters, that drops the TDP a lot so they could try cranking up the voltage again but then they don't get to save money on the heatsink being lower spec.
At some point they might check comparative performance to ATI cards, but I expect that will be secondary to just trying to get some cards out the door an generate some revenue.
Lowering the memory width doesn't make sense to me here as the main benefit of that is a cheaper PCB as it needs less traces and layers, but they already have the PCBs. Leave that trick to lower end chips.