OpenGL guy
Veteran
Certainly not pixel limited
It could be ALU/texture/bandwidth limited. There are experiments you can do to determine which it is.
NVIDIA Fermi: Architecture discussion
- Thread starter Rys
- Start date
It could be ALU/texture/bandwidth limited. There are experiments you can do to determine which it is.
Jawed
Legend
I said "pixel limited" to imply "ALU/texture/bandwidth/colour-rate/z-rate limited during pixel shading", specifically not saying "fillrate limited" as that's too specific. It's hard to split them out without having the hardware and software to hand.Certainly not pixel limited
Jawed
OpenGL guy
Veteran
Well it seems to me that by "pixel limited" you mean "hardware bound"
Ah, I see. Was only looking at their "Extreme" Scores, totally ignoring their lower-resolution runs with "extreme settings". Thanks!
Jawed
Legend
So very similar to core or shader or memory.
Compare with these:
http://www.legionhardware.com/document.php?id=869&p=9
http://www.legionhardware.com/document.php?id=869&p=10
Jawed
An underdog doesn't have more market share than its competitor
Like intel?
Is Intel the underdog now ?
Anyway NVIDIA's market share is bigger than ATI's.. NVIDIA is definitely not the underdog as someone else said a while back.
That goes without saying, since Fermi is not yet released...On the other hand, Evergreen is, but the supply is so poor that it's practically impossible to get one, without delays.
May be you should look at the "nv shows sign of strain" thread.
The losses were real [O($100M) for 2-3 (4?) quarters], although nowhere near what amd's cpu division managed to achieve.
2 quarters IIRC and they were related to the fact that they had to pay more money to cover the mobile chip failures. If it wasn't for that, they would have reported profits, during those two quarters.
That was my own speculation, without architectural differences that may improve performance. It may in fact have the same or more TMUs and the same ROPs, than the GTX 285, plus the extra shader power.
Bomba sales confirmed.
The mobile chip charges were $200M. The total losses are more than that amount.
Mintmaster
Veteran
He didn't say fillrate limited. He said pixel limited. Pixels need BW, math, filtered fetches, etc.Certainly not pixel limited
If you're not pixel limited, you have substantial parts of frame time devoted to setup or other geometry pipeline limits, vertex shading (unlikely for unified archs), CPU, or PCIe transfer. If a program doesn't scale shadow or reflection maps with resolution, then those don't scale with pixel count either.
That's what NVIDIA estimated initially, but that's not what they had to pay as a whole. In one quarter, after the problem was found they paid less than those $200 million and then in the following quarters (IIRC the next two) payed a few million more. As a whole they paid over $300 million for that. Check the quarterly results and you'll see what I mean.
Heh, well they would be quotes from several different pages and I'm on a lazy day
neliz said:
I see the HD 5870 in stock in one or two stores here. The HD 5970 is nowhere to be found. HD 4800s/GTX 200s are MIA and the HD 5700s are the only ones from the new series that I can find without problems.
