Tweet
Mehen - the 65nm process is mature enough that there really is no reason for them not to have shipped their flagship product on it. Seriously. With all the delays they had and with the architecture they were aiming at, they should have just forgotten 80nm and gone for broke w/ 65nm. The 8800 architecture is sufficiently advanced enough that even at its die size it poses serious threats.
As far as drivers... no... ATI's whole driver model is broken and has been broken for a really, really long time. The fact that they keep releasing drivers with horribly obvious problems (was it the 7.2 or the 7.3 release that people here kept having problems even installing?) is a pretty good indication that they have a really, really bad process for developing their drivers. Back when I was doing OGL programming and having to deal with the 8500/9700 shader paths (under OpenGL 1.3 using card specific render paths and GL extensions, no less) they were horrid. For every bug they fixed in a driver release, it wasn't uncommon for them to have at least one additional regression. A good driver release was only one regression.
And as far as high performance at high resolutions being proof of specific driver issue that magically cures itself at higher resolutions... bullshit. Check out AnandTech's Performance Info. Specifically, look at the graphs that show FPS versus resolution, and check each card. ATI's falls in a fairly uniform pattern - consistent with the X1950. The only cards that are having abnormal changes in performance at 2560x1600 is the 8800 series... which suggests that either 1) the GTS parts are running into the upper limits of their architectures, or 2) they have a driver problem (either bug or by design) which leads to a cumulative performance degradation at higher resolutions.
If it was a driver problem causing low performance at lower resolutions but magical performance at higher resolution the graphs would be a lot smoother across all resolutions and would show a lot less of a drop off for the 2900 as compared to the 8800. That was how I know for a fact that Matrox's Parhelia drivers had a really ****ed up OpenGL stack - The P would hold a certain plateau of performance at lower resolutions and would show little to no drop in performance when increasing resolutions.
Another thing that tends to happen with driver bugs is that one program will perform great and another will perform horribly. This is simply not the case. It performs predictably and uniformly on pretty much all benchmarks (3DMark is the only exception). It's troubling, especially when you consider that the 2600 does fairly well given its resources. My bet is that it is a design problem with their compiler/scheduler that is causing it to not scale well with the additional processing capabilities of the 2900. This should be able to be fixed with software and a lot of time, but it's not easy to do.
Or maybe not. I don't have a card here to poke at.
Best comparison I have heard right now is to the GeForce FX line. Yes, later respins and driver updates fixed a lot of the criticism people had with the product line, but it pretty much took a whole new generation to fix most of the problems. Right now, both companies are on about an 18 month gap between major product lines, with respins occurring somewhere in the middle of it. That puts ATI's respin coming out a few months before NVidia's next major release.
About the *only* thing that ATI has going for it right now is the same thing NVidia had with the GeForce FX - it gave them a much better understanding of how things worked and challenged a lot of old presumptions about what is and is not necessary with their chips. It's a new architecture, and derivative works should perform a lot faster due to tweaking the basic design. NVidia has been building their GPU's on a more modular architecture since the GeForce FX and even the 8800 series draws very heavily from the lessons learned from it. ATI will, in time, learn how to better balance their chips and software.
As far as drivers... no... ATI's whole driver model is broken and has been broken for a really, really long time. The fact that they keep releasing drivers with horribly obvious problems (was it the 7.2 or the 7.3 release that people here kept having problems even installing?) is a pretty good indication that they have a really, really bad process for developing their drivers. Back when I was doing OGL programming and having to deal with the 8500/9700 shader paths (under OpenGL 1.3 using card specific render paths and GL extensions, no less) they were horrid. For every bug they fixed in a driver release, it wasn't uncommon for them to have at least one additional regression. A good driver release was only one regression.
And as far as high performance at high resolutions being proof of specific driver issue that magically cures itself at higher resolutions... bullshit. Check out AnandTech's Performance Info. Specifically, look at the graphs that show FPS versus resolution, and check each card. ATI's falls in a fairly uniform pattern - consistent with the X1950. The only cards that are having abnormal changes in performance at 2560x1600 is the 8800 series... which suggests that either 1) the GTS parts are running into the upper limits of their architectures, or 2) they have a driver problem (either bug or by design) which leads to a cumulative performance degradation at higher resolutions.
If it was a driver problem causing low performance at lower resolutions but magical performance at higher resolution the graphs would be a lot smoother across all resolutions and would show a lot less of a drop off for the 2900 as compared to the 8800. That was how I know for a fact that Matrox's Parhelia drivers had a really ****ed up OpenGL stack - The P would hold a certain plateau of performance at lower resolutions and would show little to no drop in performance when increasing resolutions.
Another thing that tends to happen with driver bugs is that one program will perform great and another will perform horribly. This is simply not the case. It performs predictably and uniformly on pretty much all benchmarks (3DMark is the only exception). It's troubling, especially when you consider that the 2600 does fairly well given its resources. My bet is that it is a design problem with their compiler/scheduler that is causing it to not scale well with the additional processing capabilities of the 2900. This should be able to be fixed with software and a lot of time, but it's not easy to do.
Or maybe not. I don't have a card here to poke at.
Best comparison I have heard right now is to the GeForce FX line. Yes, later respins and driver updates fixed a lot of the criticism people had with the product line, but it pretty much took a whole new generation to fix most of the problems. Right now, both companies are on about an 18 month gap between major product lines, with respins occurring somewhere in the middle of it. That puts ATI's respin coming out a few months before NVidia's next major release.
About the *only* thing that ATI has going for it right now is the same thing NVidia had with the GeForce FX - it gave them a much better understanding of how things worked and challenged a lot of old presumptions about what is and is not necessary with their chips. It's a new architecture, and derivative works should perform a lot faster due to tweaking the basic design. NVidia has been building their GPU's on a more modular architecture since the GeForce FX and even the 8800 series draws very heavily from the lessons learned from it. ATI will, in time, learn how to better balance their chips and software.
Fractal Design Arc Mini R2, 3800X, Asus B450M-PRO mATX, 2x8GB B-die@3800C16, AMD Vega64, Seasonic 850W Gold, Black Ice Nemesis/Laing DDC/EKWB 240 Loop (VRM>CPU>GPU), Noctua Fans.
Comment