Rare as in, every generation? Because in case you havent noticed
6800 2x ++ 5900
7800 2x 6800
8800 2x ++ 7800
By the time it launches its been a full year since the 8800 was launched, so 2x is not surprising
rare doubling of their previous top-end
G92 will be released in November 2007 timeframe in the form of "GeForce 9800" series.
Originally posted by: chewietobbacca
Except Nvidia has never skipped to a new nm process on the high end without testing it first on the mid-range and low-end
Also, the entire memory bandwidth and memory problem is questionable considering their current G80 architecture is what they are going to be building on. IT would be weird to suddenly revise it to 512-bit / 1 gig when 384/768 and its derivatives are working.
Originally posted by: ShadowOfMyself
rare doubling of their previous top-end
Rare as in, every generation? Because in case you havent noticed
6800 2x ++ 5900
7800 2x 6800
8800 2x ++ 7800
By the time it launches its been a full year since the 8800 was launched, so 2x is not surprising
Originally posted by: Extelleron
Originally posted by: Cheex
Originally posted by: Extelleron
The 8800GTS 320MB is midrange
I'm insulted...:|
Take it back!!
If you look at the price ranges for GPUs.... from $60 to $600, $270~ is right in the middle, so technically, the 8800GTS 320 is midrange
In my opinion:
Below $100: Ultra-low end
$100-$150: Low-end
$150-$200: Lower Mid-range
$250-$300: Upper Mid-range
$350-$450: High-End
$500-$600: Ultra High-End
Overall, the 8800GTS 320MB is kind of a weird card. It's technically midrange, but really it's the same GPU as the "high-end" 8800GTS 640MB (which by extension is a hacked down version of the "ultra high-end" 8800GTX). I never understood why nVidia didn't release an 80nm, 64 shader part for the $250~ price range, rather than an extremely expensive, 90nm high-end GPU coupled with less memory than the card really needs. Memory doesn't cost much, so nVidia isn't saving much with the 320MB version. With a 480mm^2 GPU, nVidia can't be making much money selling an 8800GTS 320MB for $270~.
As for those that say 2-3x performance of G80 is not going to happen, I believe you are wrong here. My theory is that G90 will be around 2x, perhaps a bit less, the performance of G80 in DirectX 9, but closer to 3x faster in DX10. G80 is a first-generation DX10 part and I'm sure nVidia has learned more about DX10, and how developers are implementing it in games, since last year. With much more raw power AND optimizations, 3x the performance of G80 in DX10 is not that hard to imagine.
The real question is whether or not G90 will be a single die or multiple die. R700 is rumored to be multiple, and some have said G90 is as well. In my opinion, if nVidia is going for a card with approximately 2-3x the raw power of G80 (if the 1TFLOPs figure is true, G90 is actually around 3x more powerful than G80, in raw horsepower) than they need to have multiple die. For 3x the raw power, nVidia would need at LEAST 256 shaders, and they would still need to clock the shaders at well over 2000MHz. And, they would be stuck with the same old huge, expensive 480mm^2 chip they have now (with double the execution units, on a half-size process, G90 would be approximately the same size as G90). I can't see nVidia doing that.
However, if you have 2 die with, say, 160 shaders on each, then suddenly it becomes a lot more realistic. There are two die, of course, but they are smaller and the chips will have much better yields.
Originally posted by: chizow
Originally posted by: chewietobbacca
Except Nvidia has never skipped to a new nm process on the high end without testing it first on the mid-range and low-end
Also, the entire memory bandwidth and memory problem is questionable considering their current G80 architecture is what they are going to be building on. IT would be weird to suddenly revise it to 512-bit / 1 gig when 384/768 and its derivatives are working.
Aren't the 8600 and 8-series mobile parts all on 65nm? Pretty sure the latest rumors had NV scrapping the 80nm refresh (8900s) and instead choosing to go forward with a 65nm high-end part and new line. The 2900XT's performance (or lack-of) relative to the 8800GTX probably had a lot to do with that decision.
Also the current G80 memory structure is actually an odd one, with something like 5 memory controllers and 6 shader clusters making it very inflexible in terms of scaling the GPU. I forget the exact figures, but going back to a 512/1GB would be much easier for them to scale for lower-end parts.
Originally posted by: BFG10K
Going by past trends I don't think 2x performance is an unreasonable claim, especially since it's claimed it will have eDRAM.
Originally posted by: gersson
Originally posted by: BFG10K
Going by past trends I don't think 2x performance is an unreasonable claim, especially since it's claimed it will have eDRAM.
didnt ATI technically employ this first with xbox360? Ironic that nvidia might beat them to the punch on PC.
Originally posted by: coldpower27
Originally posted by: ShadowOfMyself
rare doubling of their previous top-end
Rare as in, every generation? Because in case you havent noticed
6800 2x ++ 5900
7800 2x 6800
8800 2x ++ 7800
By the time it launches its been a full year since the 8800 was launched, so 2x is not surprising
7900 GTX to 8800 GTX, achieved a good 1.6-2.0x, and that's only if you discount 7950GX2.
Originally posted by: Frackal
Originally posted by: coldpower27
Originally posted by: ShadowOfMyself
rare doubling of their previous top-end
Rare as in, every generation? Because in case you havent noticed
6800 2x ++ 5900
7800 2x 6800
8800 2x ++ 7800
By the time it launches its been a full year since the 8800 was launched, so 2x is not surprising
7900 GTX to 8800 GTX, achieved a good 1.6-2.0x, and that's only if you discount 7950GX2.
Go look at the Anand benchmarks from the original 8800 release date. Those were immature drivers (there has since been a very solid driver-optimized performance increase as seen in later tests)
The 8800GTX is ALWAYS twice as fast as the 7900GTX, and in games like Oblivion, it has 3x the framerates.
The x1900xtx is faster than the 7900GTX, I just upgraded from an X1900xtx and the 8800 is EASILY twice as fast, especially with the right CPU. It's quite incredible.
It's more like 2.0-3.0x the 7900gtx
Originally posted by: ShadowOfMyself
Not only that but the 7900 was a simple refresh, out less than 1 year from the 7800
All they did was double the memory and increase the clocks, no architectural changes (except for the shrink to 90nm)
What I mean is, we shouldnt be comparing small refreshes when talking about performance increases because obviously between those there wont be a huge jump
9700 -> 9800
X1800 -> X1900
7800 -> 7900
and so on
Originally posted by: coldpower27
Originally posted by: chizow
Originally posted by: chewietobbacca
Except Nvidia has never skipped to a new nm process on the high end without testing it first on the mid-range and low-end
Also, the entire memory bandwidth and memory problem is questionable considering their current G80 architecture is what they are going to be building on. IT would be weird to suddenly revise it to 512-bit / 1 gig when 384/768 and its derivatives are working.
Aren't the 8600 and 8-series mobile parts all on 65nm? Pretty sure the latest rumors had NV scrapping the 80nm refresh (8900s) and instead choosing to go forward with a 65nm high-end part and new line. The 2900XT's performance (or lack-of) relative to the 8800GTX probably had a lot to do with that decision.
Also the current G80 memory structure is actually an odd one, with something like 5 memory controllers and 6 shader clusters making it very inflexible in terms of scaling the GPU. I forget the exact figures, but going back to a 512/1GB would be much easier for them to scale for lower-end parts.
G80 is a 6x64Bit Configuration in full Shader Clusters should be either in Blocks of 16 or 32, as the 8800 GTS to GTX shows. so either 3/4 or 6/8
And I dunno it looks like Shader Blocks and Bit Width are pretty independent to me, as it's all on the G80 die, but your just disabling portions as you go along.
Originally posted by: manimal
You would think they would skip the 9800 number as well since it was, well, an ATI product name, it would be like starting a new car company and nameing the car the Corvair-G92,
Originally posted by: Frackal
Originally posted by: coldpower27
Originally posted by: ShadowOfMyself
rare doubling of their previous top-end
Rare as in, every generation? Because in case you havent noticed
6800 2x ++ 5900
7800 2x 6800
8800 2x ++ 7800
By the time it launches its been a full year since the 8800 was launched, so 2x is not surprising
7900 GTX to 8800 GTX, achieved a good 1.6-2.0x, and that's only if you discount 7950GX2.
Go look at the Anand benchmarks from the original 8800 release date. Those were immature drivers (there has since been a very solid driver-optimized performance increase as seen in later tests)
The 8800GTX is ALWAYS twice as fast as the 7900GTX, and in games like Oblivion, it has 3x the framerates.
The x1900xtx is faster than the 7900GTX, I just upgraded from an X1900xtx and the 8800 is EASILY twice as fast, especially with the right CPU. It's quite incredible.
It's more like 2.0-3.0x the 7900gtx
Originally posted by: coldpower27
Originally posted by: Frackal
Originally posted by: coldpower27
Originally posted by: ShadowOfMyself
rare doubling of their previous top-end
Rare as in, every generation? Because in case you havent noticed
6800 2x ++ 5900
7800 2x 6800
8800 2x ++ 7800
By the time it launches its been a full year since the 8800 was launched, so 2x is not surprising
7900 GTX to 8800 GTX, achieved a good 1.6-2.0x, and that's only if you discount 7950GX2.
Go look at the Anand benchmarks from the original 8800 release date. Those were immature drivers (there has since been a very solid driver-optimized performance increase as seen in later tests)
The 8800GTX is ALWAYS twice as fast as the 7900GTX, and in games like Oblivion, it has 3x the framerates.
The x1900xtx is faster than the 7900GTX, I just upgraded from an X1900xtx and the 8800 is EASILY twice as fast, especially with the right CPU. It's quite incredible.
It's more like 2.0-3.0x the 7900gtx
No it isn't that is only if your looking at 2560x1600 resolutions and nothing else, if you look at the review as a whole, you seen that there are times where the performance increase is only 1.6x. Simply looking at 25x16 is not representative as a whole.
Originally posted by: chizow
Originally posted by: coldpower27
Originally posted by: chizow
Originally posted by: chewietobbacca
Except Nvidia has never skipped to a new nm process on the high end without testing it first on the mid-range and low-end
Also, the entire memory bandwidth and memory problem is questionable considering their current G80 architecture is what they are going to be building on. IT would be weird to suddenly revise it to 512-bit / 1 gig when 384/768 and its derivatives are working.
Aren't the 8600 and 8-series mobile parts all on 65nm? Pretty sure the latest rumors had NV scrapping the 80nm refresh (8900s) and instead choosing to go forward with a 65nm high-end part and new line. The 2900XT's performance (or lack-of) relative to the 8800GTX probably had a lot to do with that decision.
Also the current G80 memory structure is actually an odd one, with something like 5 memory controllers and 6 shader clusters making it very inflexible in terms of scaling the GPU. I forget the exact figures, but going back to a 512/1GB would be much easier for them to scale for lower-end parts.
G80 is a 6x64Bit Configuration in full Shader Clusters should be either in Blocks of 16 or 32, as the 8800 GTS to GTX shows. so either 3/4 or 6/8
And I dunno it looks like Shader Blocks and Bit Width are pretty independent to me, as it's all on the G80 die, but your just disabling portions as you go along.
Right, but I remember there was some pretty lengthy discussion about the memory configuration of the G80 when the 320MB part launched and that the traces for the memory interface showed an imbalance, with some shader quads being serviced by a disproportionate number of memory interfaces/modules. I think AT even mentioned this as a road block as to why they couldn't put more memory on the 320MB model.
Ideally you'd want everything to be balanced and proportionate, which isn't the case with the GTS and GTX. If you went to an 8x64 bit configuration, you'd have 2:1 memory/shader ratio for the GTX or 1:1 ratio if you doubled the shaders on a G9X part. If you wanted to trim the part down for a mid-range model, you could simply cut the shaders and memory interfaces in half and still maintain a 1:1 ratio.