Isn't it wonderful how, in this modern day and age, that the first and seemingly most popular approach taken by establishment against innovation is a law suit?
Wow, that sounds disturbingly like something Micro$oft would say to the DOJ...
Who said that writing drivers are resources splitting?
I'm almost sure that the Linux drivers are done by Precision Insight (including Itanium), and I don't think that the Mac developers help the Windows drivers developers at all..
They still have to pay them, don't they? It all comes out of the driver budget. So more driver platforms still means less money per platform, which means less time spent on optimisation, regardless of whether it's done in-house or not.
Namarrgon
Re:Going through the features...
on
ATI Radeon 256
·
· Score: 1
No, the memory runs at 200MHz, it doesn't mention that the chip itself will run at 200MHz.
In fact, they don't even claim that, only that the memory COULD run at 200MHz. I haven't heard of 200MHz DDR RAM, but I bet if it exists, it's expensive.
this chip could be running at 300-400MHz
What I've read suggests that the first chip will actually run at 200MHz. Faster clock speeds may come in the future, perhaps after another die shrink.
But 1500MTexels/3 = 500M Pixels per second. If the chip can effectively commit 1 pixel per clock cycle that would be 500MHz, I suspect it can commit 2 pixels per clock cycle, or even more!
If the clock speed of the chip is 200MHz as FiringSquad suggest, then the real max fillrate is 1.2 MTexels/s, discounting HYPErZ (which no-one has explained to my satisfaction yet).
For older, single-textured games, that's 400-500 MPixels/s, about GeForce speed. For the current crop of dual-textured games, it should be up to twice as fast, (memory bandwidth permitting, of course). If we ever see games with 3 textures per pixel, then they might be able to get within cooee of the press release..
Contrast this to the new 3dfx V5-5500 (667 MTexel/s, faster than Radeon for single-textured games, slower for dual or more), or the about-to-be-announced GeForce2, which will reportedly do 800 MPixels/s, single OR dual textured (i.e. up to 1600 MTexels/s). IF the DDR memory can keep up (ho ho).
It'd help a lot more if there was an HDTV tuner/receiver onboard. How much are they going to cost?
Being able to decode HDTV-size MPEG2 is no bad thing, but not particularly exciting in itself. With a cheap receiver (not to mention some actual content), it'd be more interesting.
And MAXX'ed? Didn't we already see that MAXX technology does undesireable things to one's latency? Who cares if you get 150fps if you're a half-step behind everyone else?
Well, with 1 frame latency at 150 fps, that's only 6.7 milliseconds - not much of a step behind. Perhaps at 10 fps it'd be a problem.
OTOH, drivers have always been ATI's weak point. Their last card had drivers for Win9x ONLY (not even NT), and even those got dragged down by excessive CPU usage. It's all very well them announcing Linux, Mac and even Itanium drivers, but that's just going to split their driver resources even further. Who thinks developing drivers for more platforms is going to result in better drivers for any platform?
Their hardware sounds cool, as cool as the GeForce2, but that's no use if the drivers suck. Perhaps they should just do the hardware, open the driver source, and let everyone else do all the different platforms for them.
But the human eye can tell the difference between 30 and 60 fps. Look closely at movie with lots of action and you will notice the individual frames. That is at 24 fps but US television at 30 fps would appear just as choppy if the resolution were higher.
US television (NTSC) is actually 60 fields per second - with each successive field interlaced to provide a full resolution frame, but 60 Hz nonetheless. And movies are shown at 72 Hz, not 48 (which would still flicker too much).
It's quite easy to tell the difference between between 30 fps and 60 fps. It's also possible to tell the difference between 60 fps and 75 fps - have a look at a computer screen set to 60 Hz refresh rate, then set it to 75 Hz. 60 Hz is annoyingly flickery.
I believe video cards will continue to develop long past the point of 75 Hz @ 1600 x 1200, or even at higher resolutions. Once sufficient speed at the best res current monitors can do is attained, greater and greater speed will be needed for better full-screen antialiasing instead. But there are huge advances still needed in quality.
When you compare Q3A or UT against Toy Story, you can see what they're aiming at, and how far they have to go. Then compare Toy Story to The Matrix, The Mummy, or Episode 1. Finally, look around - reality itself is the ultimate target.
Recorded audio reproduction has already reached the point where realism is only an issue with purists. Dynamically generated audio isn't doing too badly either, though it doesn't have the dollars behind it that video does. Video has far more to live up to, to fool human eyes and brains. Believe me, we won't be seeing a slowdown there anytime soon.
Dunno bout benchmarking, but as I understand it, neither windows nor linux can claim full CPU usage in a SMP system. So it seems it would be fairly much impossible for either of them to even compete with Be, which uses 100%.
I think this is simply because no one makes SMP apps for linux or windows. There is just no demand. So the OS will have to try and make up for that with all these apps that aren't even thread safe.
Sorry, I had to laugh when I read this:-)
As the developer of a very threaded SMP app (for post production) for Windows NT/2000 (and competing with numerous other SMP apps), I can assure you this is not at all the case. Win9x may be single CPU, but there is plenty of demand for raw speed out there, and SMP is still the cheapest way to get it.
Our app not only uses 100% of all CPUs, even on an 8-CPU system, it achieves over 80% of linear performance scaling while doing it. Over 90% for dual and quad systems (the 8-ways have some architectural peculiarities). It's quite possible under Windows, despite the myriad annoyances of the OS.
Be's good/cheap SMP was the main reason I bought a BeBox so long ago, and why I still have a fond eye for it (that, and its API). HOWEVER:
We ran some earlier tests (a year a two ago) on efficiency of the kernal's thread-switching, and a little to our surprise, we found NT's kernal was slightly (1% or so) more efficient than BeOS's, when running executables on each compiled by the same (MS) compiler, and linked by the native linker (I'm not sure if this is possible any longer, due to subsequent changes by Be).
OTOH, when testing Be's native compiler vs. the Intel Windows reference compiler, there were huge differences (30%-80% faster(!) under NT, even more for some operations).
This is great for computationally-intensive Windows programs (like ours), but during interactive usage, all of this gain is swallowed by Window's clunky, not-very-threaded interface (and don't even get me started on the abomination that is MFC). So BeOS does often "feel" faster and slicker.
Seems like a lot of people are missing the point here. It's not designed for running Linux or Office, but for playing games. It won't appeal to PC gamers (they've already got PCs), it'll appeal to console gamers - and there's a lot of them about. It'll offer the price, and all the power-up-&-go convenience of a console, with the ease of development of a PC.
Some specs (e.g. CPU, RAM, hard drive) will be low-end by general-purpose PC standards, when it's released, but the graphics will be state-of-the-art. Sure your new PC will run POVray twice as fast, but I bet the X-Box games will be similar, or better. Developers can usually take better advantage of a known, stable platform than they can with a PC, because they don't have to cater for the lowest common denominator.
Although it'll be running "a version of Windows 2000", you won't see it. There won't be configuration problems, or patches to install, you just pop in your game CD & go. The OS will boot fast because it'll simply load a RAM image from the HD, a la Win2K hibernation. They're going to hide the PC-ness from the user, and present only the console face.
Of course, clever people will doubtless find a way to install Linux on the hard drive, and then suddenly they'll have a $200 PC with networking, sound, fast gfx and a DVD drive, but to the 71 million PlayStation owners out there, it'll be a new console they can run their PSX games on (via Bleem!), but which is just as easy to use as their PSX2, and looks considerably nicer. Not to mention that it'll get all the latest games a lot sooner than the PSX2 will, due to the ease of porting from the PC (ever heard how hard it is to use those wonderful PSX2 features?).
I think the hardware will certainly be competitive (with other consoles), and the games will be plentiful (so long as MS get their marketing right, and they've started off on the right foot - the PC developers looked happy). MS say the price will be comparable to other consoles ($150-$300).
So, think of it as a console, with easy to use, consistant games, with the potential to be a really cheap, basic PC. Just don't expect anyone to replace their main PC with one.
It's just like Deep Blue vs. Kasparov again. Brute force compilers vs. intuitive & creative humans. In some cases, humans can write better assembly than compilers (thanks to knowing more about what the program is intended to do), and in others the compilers have the advantage (thanks to being better at scheduling instructions on today's pipelined, superscalar CPUs). JIT compilers could have even more knowledge about instruction timing, but have less time to think about it. Kinda like Deep Blue playing blitz chess. I'd say in most cases these days, a compiler could write faster code from scratch than a human could, unless the human really spends some time on it, but the optimal combination will always be a human assisted by a compiler (and profiler, etc). Best of both worlds. Namarrgon
The driver didn't access all the features of the chipset. There was no advantage in having a TNT2 with 32Mb and a TNT with 16Mb. (Unless you had a ridiculously large Workspace with 32-bit colour, but that has nothing to do with the GLX code)
I have a secret for you: no one cares. 32 megs is on that card 'cos they wanted to have a bigger number on there, but there's no excuse for it. The only case in which you would care is if you literally have 32 megs of textures on the screen at once. Otherwise you can page on and off the card very efficiently. Speaking from experience here, about 4 megs of texture memory is more than plenty.
Actually, 32Mb of texture RAM *does* make a big difference, particularly with the newer games. 8-10 months ago, when 16 Mb cards were the norm and the first 32 Mb cards came out, numerous benchmarks showed the bigger RAM helped significantly when moving to 32 bit display & textures.
Yes of course you can swap textures in & out, but it's hardly free - bus bandwidth is one of the most stressed resources in today's games (hence the popularity of AGP). And it can slow things down dramatically. Q3Test takes quite a decent performance hit on my 16 MB TNT when you turn on 32 bit textures (doubling the texture RAM needed). AGP texturing can help, but is still slower than more local RAM.
Finally, don't forget this RAM is usually also used for the frame buffer and z buffer - at 1280x1024x32 for each you've already used up over 10 MB just for those.
Besides, by the time this card actually gets realeased, I'll be running dual 1 Ghz athalon's on my DDR SDRAM board hopefully, and will just need darn fast fillrate;-) My processors will be faster then hardware t&l for the near future I think, in another year or two there should be a card with decent fillrate and fast T&L that I will buy...
I doubt that. OpenGL can only take advantage of one of your processors, and only one game (Q3A) makes any attempt to use the second CPU in any other way.
And the GeForce T&L is *way* faster than anything else out there, up to the level of the SGI InfiniteReality2. Check out the Indy3D and viewperf benchmarks - the GeForce pushes more textured smoothed polys than anything else. Why do you think professionals spend $3k or more on a card with T&L a fifth the speed of the GeForce? Because even that is so much faster than the CPU, especially once you include physics, AI etc in the CPU load.
The chips all work in parallel, rendering strips of the frame, so the fillrate will scale on any app. They're emphasizing the pixels per clock partly because nVidia are touting their QuadEngine(tm) that does 4 pixels/clock, and partly because their rendering pipeline has changed.
There is a difference between the V2/V3 and V4/V5. When rendering single-textured pixels, one texture unit on the older cards sat idle, and they didn't really come into their own until the app did multitexturing. The new chips have switched to a dual pixel pipeline, similar to the TNT/TNT2, which allows two single textured pixels or (presumably) one dual textured pixel per clock. (Note that "fully-featured" does not mean "dual-textured", though they seem to want to imply that.) So their performance will halve when dual-texturing.
Unfortunately (and no-one seems to have considered this), the memory is split between the chips too. This means the highly expensive V5 6000 with 128MB of RAM can really only use 32MB of RAM, and the V5 5000 can only use 16MB, which will certainly restrict its performance.
Something else no-one has talked about anywhere is how seriously the anti-aliasing will slow things down. 3dfx have said there will be a performance drop, but skirted around how much. The fact is, it'll slash your fillrate by 75%! So your $300 V5 5500, with two chips and 64MB of RAM will perform at less than the speed of a 1998 TNT when you flick that switch. Sure, it'd look nicer, but I'd rather run faster at a higher resolution without AA - the edges look less jagged and you get more detail as well.
As for the V4.. it will offer the same performance & features as a TNT2 Ultra does, only a year later and for over $50 more. I can't see anyone buying that at all.
3dfx may regain the performance crown (for a month, until/if the Glaze3D comes out), but they're still 2 years behind in rendering quality. A year from now, hardware T&L will be the norm on all current chips, except 3dfx's. Glad I'm not a 3dfx shareholder (did you know their stock dipped $5 on the day they announced their new line? Ouch.)
Seriously, though, sheer pixels per second is kind-of meaningless. (Actually, it'll be state changes per second.:) In most vector-based or polygon-based product, the bottle-neck is in calculating the outer perimiter of the shapes.
Not always; depends what you're doing. With simple scenes (less than 10k polygons), and using texture maps, lightmaps, bump maps, translucency maps etc, at high resolutions (if you want to reduce aliasing), major pixel fillrate is needed - it's still a big bottleneck (one of many).
Personally, I think pixel-based displays are a dead-end, anyway. Aliasing is horrible, and the techniques to get round it do so by making the picture too blurred to tell.
Reality itself is aliased, it's just that the pixels are really really small. There's no such thing as pure analog. And anti-aliasing techniques do not soften or blur the picture at all (like filtering techniques do), they make the edges cleaner, but still crisp. Supersampling (a popular method of anti-aliasing that requires huge fillrates) duplicates what our vision does (the eye "supersamples" what it sees, by averaging all the rays that fall on a particular rod or cone).
We don't have the retinal resolution to perceive the world around us at the atomic (let alone the Planck) level, so we see those nasty jagged edges as perfectly smooth. Once pixel-based displays get close to the limits of our eyes (1200dpi looks pretty good), we won't be able to tell the difference.
Not to say there aren't more efficient ways of doing this, but pixel displays will work perfectly well.
Slashdot Mirror
We love loss leaders. They give us cheap hardware to play with.
You bet. Just wait till I get my 16 x XBox renderfarm set up...
Who said that writing drivers are resources splitting?
I'm almost sure that the Linux drivers are done by Precision Insight (including Itanium), and I don't think that the Mac developers help the Windows drivers developers at all..
They still have to pay them, don't they? It all comes out of the driver budget. So more driver platforms still means less money per platform, which means less time spent on optimisation, regardless of whether it's done in-house or not.
Namarrgon
No, the memory runs at 200MHz, it doesn't mention that the chip itself will run at 200MHz.
In fact, they don't even claim that, only that the memory COULD run at 200MHz. I haven't heard of 200MHz DDR RAM, but I bet if it exists, it's expensive.
this chip could be running at 300-400MHz
What I've read suggests that the first chip will actually run at 200MHz. Faster clock speeds may come in the future, perhaps after another die shrink.
But 1500MTexels/3 = 500M Pixels per second. If the chip can effectively commit 1 pixel per clock cycle that would be 500MHz, I suspect it can commit 2 pixels per clock cycle, or even more!
If the clock speed of the chip is 200MHz as FiringSquad suggest, then the real max fillrate is 1.2 MTexels/s, discounting HYPErZ (which no-one has explained to my satisfaction yet).
For older, single-textured games, that's 400-500 MPixels/s, about GeForce speed. For the current crop of dual-textured games, it should be up to twice as fast, (memory bandwidth permitting, of course). If we ever see games with 3 textures per pixel, then they might be able to get within cooee of the press release..
Contrast this to the new 3dfx V5-5500 (667 MTexel/s, faster than Radeon for single-textured games, slower for dual or more), or the about-to-be-announced GeForce2, which will reportedly do 800 MPixels/s, single OR dual textured (i.e. up to 1600 MTexels/s). IF the DDR memory can keep up (ho ho).
Namarrgon
It'd help a lot more if there was an HDTV tuner/receiver onboard. How much are they going to cost?
Being able to decode HDTV-size MPEG2 is no bad thing, but not particularly exciting in itself. With a cheap receiver (not to mention some actual content), it'd be more interesting.
Namarrgon
their drivers are somewhat less than stellar.
And MAXX'ed? Didn't we already see that MAXX technology does undesireable things to one's latency? Who cares if you get 150fps if you're a half-step behind everyone else?
Well, with 1 frame latency at 150 fps, that's only 6.7 milliseconds - not much of a step behind. Perhaps at 10 fps it'd be a problem.
OTOH, drivers have always been ATI's weak point. Their last card had drivers for Win9x ONLY (not even NT), and even those got dragged down by excessive CPU usage. It's all very well them announcing Linux, Mac and even Itanium drivers, but that's just going to split their driver resources even further. Who thinks developing drivers for more platforms is going to result in better drivers for any platform?
Their hardware sounds cool, as cool as the GeForce2, but that's no use if the drivers suck. Perhaps they should just do the hardware, open the driver source, and let everyone else do all the different platforms for them.
Namarrgon
But the human eye can tell the difference between 30 and 60 fps. Look closely at movie with lots of action and you will notice the individual frames. That is at 24 fps but US television at 30 fps would appear just as choppy if the resolution were higher.
US television (NTSC) is actually 60 fields per second - with each successive field interlaced to provide a full resolution frame, but 60 Hz nonetheless. And movies are shown at 72 Hz, not 48 (which would still flicker too much).
It's quite easy to tell the difference between between 30 fps and 60 fps. It's also possible to tell the difference between 60 fps and 75 fps - have a look at a computer screen set to 60 Hz refresh rate, then set it to 75 Hz. 60 Hz is annoyingly flickery.
I believe video cards will continue to develop long past the point of 75 Hz @ 1600 x 1200, or even at higher resolutions. Once sufficient speed at the best res current monitors can do is attained, greater and greater speed will be needed for better full-screen antialiasing instead. But there are huge advances still needed in quality.
When you compare Q3A or UT against Toy Story, you can see what they're aiming at, and how far they have to go. Then compare Toy Story to The Matrix, The Mummy, or Episode 1. Finally, look around - reality itself is the ultimate target.
Recorded audio reproduction has already reached the point where realism is only an issue with purists. Dynamically generated audio isn't doing too badly either, though it doesn't have the dollars behind it that video does. Video has far more to live up to, to fool human eyes and brains. Believe me, we won't be seeing a slowdown there anytime soon.
Namarrgon
As the developer of a very threaded SMP app (for post production) for Windows NT/2000 (and competing with numerous other SMP apps), I can assure you this is not at all the case. Win9x may be single CPU, but there is plenty of demand for raw speed out there, and SMP is still the cheapest way to get it.
Our app not only uses 100% of all CPUs, even on an 8-CPU system, it achieves over 80% of linear performance scaling while doing it. Over 90% for dual and quad systems (the 8-ways have some architectural peculiarities). It's quite possible under Windows, despite the myriad annoyances of the OS.
Be's good/cheap SMP was the main reason I bought a BeBox so long ago, and why I still have a fond eye for it (that, and its API). HOWEVER:
We ran some earlier tests (a year a two ago) on efficiency of the kernal's thread-switching, and a little to our surprise, we found NT's kernal was slightly (1% or so) more efficient than BeOS's, when running executables on each compiled by the same (MS) compiler, and linked by the native linker (I'm not sure if this is possible any longer, due to subsequent changes by Be).
OTOH, when testing Be's native compiler vs. the Intel Windows reference compiler, there were huge differences (30%-80% faster(!) under NT, even more for some operations).
This is great for computationally-intensive Windows programs (like ours), but during interactive usage, all of this gain is swallowed by Window's clunky, not-very-threaded interface (and don't even get me started on the abomination that is MFC). So BeOS does often "feel" faster and slicker.
Daniel Koch
daniel@eyeonline.com
Seems like a lot of people are missing the point here. It's not designed for running Linux or Office, but for playing games. It won't appeal to PC gamers (they've already got PCs), it'll appeal to console gamers - and there's a lot of them about. It'll offer the price, and all the power-up-&-go convenience of a console, with the ease of development of a PC.
Some specs (e.g. CPU, RAM, hard drive) will be low-end by general-purpose PC standards, when it's released, but the graphics will be state-of-the-art. Sure your new PC will run POVray twice as fast, but I bet the X-Box games will be similar, or better. Developers can usually take better advantage of a known, stable platform than they can with a PC, because they don't have to cater for the lowest common denominator.
Although it'll be running "a version of Windows 2000", you won't see it. There won't be configuration problems, or patches to install, you just pop in your game CD & go. The OS will boot fast because it'll simply load a RAM image from the HD, a la Win2K hibernation. They're going to hide the PC-ness from the user, and present only the console face.
Of course, clever people will doubtless find a way to install Linux on the hard drive, and then suddenly they'll have a $200 PC with networking, sound, fast gfx and a DVD drive, but to the 71 million PlayStation owners out there, it'll be a new console they can run their PSX games on (via Bleem!), but which is just as easy to use as their PSX2, and looks considerably nicer. Not to mention that it'll get all the latest games a lot sooner than the PSX2 will, due to the ease of porting from the PC (ever heard how hard it is to use those wonderful PSX2 features?).
I think the hardware will certainly be competitive (with other consoles), and the games will be plentiful (so long as MS get their marketing right, and they've started off on the right foot - the PC developers looked happy). MS say the price will be comparable to other consoles ($150-$300).
So, think of it as a console, with easy to use, consistant games, with the potential to be a really cheap, basic PC. Just don't expect anyone to replace their main PC with one.
Namarrgon
It's just like Deep Blue vs. Kasparov again. Brute force compilers vs. intuitive & creative humans. In some cases, humans can write better assembly than compilers (thanks to knowing more about what the program is intended to do), and in others the compilers have the advantage (thanks to being better at scheduling instructions on today's pipelined, superscalar CPUs). JIT compilers could have even more knowledge about instruction timing, but have less time to think about it. Kinda like Deep Blue playing blitz chess. I'd say in most cases these days, a compiler could write faster code from scratch than a human could, unless the human really spends some time on it, but the optimal combination will always be a human assisted by a compiler (and profiler, etc). Best of both worlds. Namarrgon
Yes of course you can swap textures in & out, but it's hardly free - bus bandwidth is one of the most stressed resources in today's games (hence the popularity of AGP). And it can slow things down dramatically. Q3Test takes quite a decent performance hit on my 16 MB TNT when you turn on 32 bit textures (doubling the texture RAM needed). AGP texturing can help, but is still slower than more local RAM.
Finally, don't forget this RAM is usually also used for the frame buffer and z buffer - at 1280x1024x32 for each you've already used up over 10 MB just for those.
Namarrgon
Besides, by the time this card actually gets realeased, I'll be running dual 1 Ghz athalon's on my DDR SDRAM board hopefully, and will just need darn fast fillrate ;-) My processors will be faster then hardware t&l for the near future I think, in another year or two there should be a card with decent fillrate and fast T&L that I will buy...
I doubt that. OpenGL can only take advantage of one of your processors, and only one game (Q3A) makes any attempt to use the second CPU in any other way.
And the GeForce T&L is *way* faster than anything else out there, up to the level of the SGI InfiniteReality2. Check out the Indy3D and viewperf benchmarks - the GeForce pushes more textured smoothed polys than anything else. Why do you think professionals spend $3k or more on a card with T&L a fifth the speed of the GeForce? Because even that is so much faster than the CPU, especially once you include physics, AI etc in the CPU load.
The chips all work in parallel, rendering strips of the frame, so the fillrate will scale on any app. They're emphasizing the pixels per clock partly because nVidia are touting their QuadEngine(tm) that does 4 pixels/clock, and partly because their rendering pipeline has changed.
There is a difference between the V2/V3 and V4/V5. When rendering single-textured pixels, one texture unit on the older cards sat idle, and they didn't really come into their own until the app did multitexturing. The new chips have switched to a dual pixel pipeline, similar to the TNT/TNT2, which allows two single textured pixels or (presumably) one dual textured pixel per clock. (Note that "fully-featured" does not mean "dual-textured", though they seem to want to imply that.) So their performance will halve when dual-texturing.
Unfortunately (and no-one seems to have considered this), the memory is split between the chips too. This means the highly expensive V5 6000 with 128MB of RAM can really only use 32MB of RAM, and the V5 5000 can only use 16MB, which will certainly restrict its performance.
Something else no-one has talked about anywhere is how seriously the anti-aliasing will slow things down. 3dfx have said there will be a performance drop, but skirted around how much. The fact is, it'll slash your fillrate by 75%! So your $300 V5 5500, with two chips and 64MB of RAM will perform at less than the speed of a 1998 TNT when you flick that switch. Sure, it'd look nicer, but I'd rather run faster at a higher resolution without AA - the edges look less jagged and you get more detail as well.
As for the V4.. it will offer the same performance & features as a TNT2 Ultra does, only a year later and for over $50 more. I can't see anyone buying that at all.
3dfx may regain the performance crown (for a month, until/if the Glaze3D comes out), but they're still 2 years behind in rendering quality. A year from now, hardware T&L will be the norm on all current chips, except 3dfx's. Glad I'm not a 3dfx shareholder (did you know their stock dipped $5 on the day they announced their new line? Ouch.)
Namarrgon
(P.S. Seen this ad for a Voodoo6 8000?)
Seriously, though, sheer pixels per second is kind-of meaningless. (Actually, it'll be state changes per second. :) In most vector-based or polygon-based product, the bottle-neck is in calculating the outer perimiter of the shapes.
Not always; depends what you're doing. With simple scenes (less than 10k polygons), and using texture maps, lightmaps, bump maps, translucency maps etc, at high resolutions (if you want to reduce aliasing), major pixel fillrate is needed - it's still a big bottleneck (one of many).
Personally, I think pixel-based displays are a dead-end, anyway. Aliasing is horrible, and the techniques to get round it do so by making the picture too blurred to tell.
Reality itself is aliased, it's just that the pixels are really really small. There's no such thing as pure analog. And anti-aliasing techniques do not soften or blur the picture at all (like filtering techniques do), they make the edges cleaner, but still crisp. Supersampling (a popular method of anti-aliasing that requires huge fillrates) duplicates what our vision does (the eye "supersamples" what it sees, by averaging all the rays that fall on a particular rod or cone).
We don't have the retinal resolution to perceive the world around us at the atomic (let alone the Planck) level, so we see those nasty jagged edges as perfectly smooth. Once pixel-based displays get close to the limits of our eyes (1200dpi looks pretty good), we won't be able to tell the difference.
Not to say there aren't more efficient ways of doing this, but pixel displays will work perfectly well.