Standard short-sighted view. People said the same about ATi not long ago, and there was a lot less reason to think they'd ever turn themselves around like they did. Gamers switched to ATi's camp pretty quickly; they'll switch back even quicker if nVidia pull ahead again.
Besides, although the FX Ultra may not sell well, the effort was definitely valuable. The Quadro FX (based on FX 5800 non-Ultra) is the easily the fastest professional card out there - it destroys the ATi FireGL X1 (9700 PRO-based) in speed, quality, precision and programmability (and yes, it's quiet). The standard 5800 also has a fair chance of success, if they can price it against the ATi 9500 PRO.
Also, nVidia have doubtless learned a lot about the 0.13um process, whereas ATi have yet to jump that hurdle (R350 will still be 0.15um). Rather than being behind, nVidia might just have a headstart...
The original nForce was more interesting than successful, but clearly you haven't been keeping abreast of nForce2. Every single review has been praising it - it has virtually every feature under the sun, including some no other manufacturer has (dual DDR on AMD, Dolby Digital encoding), and its performance blows everything else away. They're walking off the shelves.
3dfx alienated their OEM customers, and had no other income apart from their consumer gfx cards. When that was delayed & out-competed (again), it was inevitable that they'd go under.
nVidia are a larger company with a string of huge successes to date. They have a much more diversified income, including some very popular OEM chips, the successful nForce2 (and less-successful Xbox) chipsets, a well-regarded pro card line, and a significant share of the Apple market too. Not to mention quite a bit of cash in the bank.
A single high-end chip(which is a small % of their total revenue anyway), even if it failed completely, is not going to impact their bottom line that much. It'll have more impact on their image as graphics leader, but they have the resources to learn, move on, redesign and try again.
All the other systems have Dolby 5.1 capability, it just depends on if the games support it.
Completely false. The GameCube doesn't even have a digital audio out. The best it can do is ProLogic II. The PS/2 can do pre-encoded Dolby Digital 5.1, good for cutscenes and DVDs, but nothing in-game. A (very) few games do in-game encoding to DTS, but this requires dedicating one of the two vector processors to sound encoding.
With the Xbox, the Dolby 5.1 encoding is done by two dedicated DSPs on the sound chip, and is automatic for all games. It's the only one with the grunt (and the licence from Dolby) to do it.
The network adapter and built-in hard drive don't mean swat because nobody will be taking full advantage of them until the other two consoles get them.
There are ports of PS/2 games that have had custom soundtrack playback added, from the hard drive. There are a number of games that require the hard drive on the Xbox, Morrowind and Project Ego being a couple - PC ports may well require it. And of course, there are still Xbox exclusives that use it to good and occasionally innovative effect, such as Blinx.
XBox Live is making very good use of the ethernet port. So are tunneling apps like XBConnect that let you play peer-to-peer online multiplayer Halo/Mech Assault/Tony Hawk/Unreal/whatever. Can't do that with just an iLink connector.
Hell, most developers still make XBox games that look like ass and by the time they figure out how to take advantage of that hardware, the PS3 will be out.
And I suppose developers will instantly know how to use the (considerably more complex) PS3 hardware to best effect? Quite the opposite - look at how long it took before the PS2 hardware got used properly. Xbox has a big advantage here, with its far easier/more familiar dev environment.
Most PS2 games look like ass too. It depends on the quality of the game, of course. Halo is very nice looking, Rallisport Challenge looks superb, Splinter Cell is amazing and Halo2 is looking awesome. PS3 is still a while away, and Xbox2 is also scheduled for that timeframe.
Actually, nVidia have been doing this for a number of years too. Full stereo drivers (not some kind of hack) are available from their site that offer a lot of flexibility (e.g. different stereo methods, per-app convergence settings etc). The cards don't need a dedicated port, as one can simply use the vsync signal to flip sides - this works with pro equipment right down to $50 consumer LCD-shutter "toys".
Speed isn't an issue either - the cheapest GeForce MX card can do 1280x1024x32 at 150Hz (75Hz left & right signals) and can still pump out triangles & pixels a lot faster than the old Creator3D.
There were 3dfx-based consumer stereo setups in 1999, and I believe nVidia introduced their stereo drivers in 2000. Cards & cheap LCD-shutter glasses have been around for a long time; they were popular for a while as a novelty for playing 3D games.
Thing is, Sun equipment is expensive; it uses a number of dedicated ASICs and has a limited market, and is priced accordingly. The money isn't there for serious development. Consumer 3D cards generate far more revenue, and have therefore progressed amazingly in the last four years, to the point where ATI's and nVidia's consumer cards are now being sold (slightly tweaked) to professionals, easily exceeding all other professional cards in speed and features, without sacrificing quality.
Thousands of layers? The most complex composition I've seen personally was the Swordfish Ventura Bank explosion, and that required somewhat over 500 layers (at 4K). Definitely qualifies as heavy duty, but still a far cry from "thousands" of layers. "Often" would be less than 100 layers, in my experience.
Anyway, tiling as you describe is rarely used in motion picture image processing work, regardless of the number of layers. Breaking down a large (4000x3000 or larger) image does improve memory usage (sometimes at a cost in efficiency for certain algorithms), but when this is done, it's usually broken into scanlines or groups of scanlines, not square tiles. This works just as well and fits better with how images are processed, stored, displayed etc. The number of layers to be composited does not affect this at all.
DPX and Cineon do not support tiled image packing. TIFF does, but I've never seen a post-production app actually output a tiled image - it just complicates things unnecessarily.
And it's rarely necessary to re-read an entire image if you just want a subrectangle of it - many formats make it relatively easy to read a limited region. Compression can complicate things, but you can usually limit your reading to just the scanlines involved.
All film cameras (still and moving) record some high-dynamic range information, yes, and the better digital cameras do too. Video cameras don't, however, and couldn't store it if they did.
Most 3D packages render directly to high-dynamic range formats, including (I think) Povray. You don't need to render multiple exposures.
16 bit float is just one of the datatypes it supports. The particular format they chose is not limited to Cg or the GeForceFX, it's the most common 16 bit float format out there, even if it isn't an IEEE standard.
The DirectX HLSL is (deliberately) syntacticly identical to Cg, so that's actually good for Cg, rather than killing it off as you suggest. OpenGL2's HLSL has yet to be confirmed, but if (as may be likely) it isn't also just like Cg, Cg will still be able to compile to OpenGL2 - it's just another render target, along with DX8/9, OpenGL 1.3/1.4, and nVidia's own extensions.
HDR info is useful for many many things in both 3D and 2D work (though I'm doubtful about edge detection). Other HDR-supporting formats do exist, including HDRI, TIFF, FLX, and RLA. Even Cineon/DPX supports limited HDR info. Each have their own advantages & disadvantages - OpenEXR is no worse than most, and better than many.
Rendering technique is only one small part of the whole job. If you want to take HDR info from one device/app/system to another, you have to write it into a file, so you need a file format that won't clip all your highlights...
ATI's "next-generation" chip is still built on the previous-generation process - 0.15 micron.
It was nVidia's move to 0.13 micron that delayed the GeForceFX, and allowed ATI their moment in the sun. ATI have yet to climb that particular hill, and nVidia are already rolling down the far side.
At the University, we were using Creator3D graphics cards from Sun Microsystems. That was in 1999, and the general consumer market still hasn't caught up with that tech.
Is there something magic about the Creator3D cards that I don't know about? According to Sun, they could do 3D at 1280x1024, had 15MB of framebuffer RAM, and could only do 1.4M tri/sec. There's no indication that it even supported texture mapping.
At first glance, the consumer market had already caught up in 1999 - the GeForceDDR comfortably exceeded those specs. These days, an $80 gfx in your Dad's new PC could probably beat it out. Heck, even the integrated gfx in an nForce2 board has higher specs.
There's nothing specific to Windows that prevents multiple capture cards. I can choose to capture from my ATI AIW's S-Video input, composite input, TV tuner or Firewire port, or my nForce2 mobo's Firewire port.
The new AIW also allows plugging in an additional PCI capture card (TV Wonder) and getting PiP functionality, or watch one/record another.
The Inquirer have "clarified" their earlier statement:
Clarification: AMD asks us to point out that Hammer schedules haven't slipped from its previous advice, as we originally suggested in this article. A spokesman from the company told us that desktop versions of Hammer are still planned to ship (for revenue) in Q1 2003 with systems on shelves at the turn of Q1 2003, not the second half of 2003 as we stated.
Fair enough, though I'm not at all convinced that many film stocks could resolve 0.005 mm lines either. Perhaps a low-ISO, ultra-low grain stock could.
While an 11 Mpixel sensor does have 11 million spatially separate elements, they are monochrome, filtered to capture R-G-B or C-M-Y, and cross-interpolated to provide a true colour image. This gets you close to 11 Mpixels of luminance resolution, but almost a third of that in chroma resolution. A RAW format image from such a sensor could probably be exported to a pretty sharp B&W image though, good enough to resolve alternating black & white lines.
There's also the Sigma SD9, using a Foveon X3 sensor. Only 3.4 Mpixels, but with RGB elements for each one, stacked vertically. That means the chroma resolution is almost that of the Canon 1Ds, but without any artifacting caused by the cross-interpolation of the colour elements. It more closely mimics film. Can't wait to see what a 14 Mpixel Foveon sensor can do:-)
a) at 100 pixels/mm, a 24 x 35mm image is only 8.4 Mpixel, not 34M.
b) You won't get anything usable from scanning 35 mm film beyond around 4000 x 3000 anyway, with most film stocks - the grain overwhelms the pixel size.
c) The Canon 1Ds (and Kodak 14n) have 12 bit sensors, which gives a dynamic range of 1:4096.
d) The Kodak DCS 14n is built with a standard Nikon SLR lens mount. The Canon EOS 1Ds is compatible with over 60 of Canon's EF lenses.
OK, a decent SLR is a lot cheaper, but it doesn't have any of the advantages a digital camera gives you, either.
Experiences differ, I suppose. I've had far more response & help from nVidia than from ATI.
With ATI, they answered my questions accurately but briefly, and made no particular response to the bug reports I posted (which finally seem to be fixed now - mostly - in 02.3, 8 months after I notified them). Their dev support team gave me adequate support, IMHO. Their regular user technical support was not even close to adequate, the once or twice I've used them - I had to track down the fix myself and explain it to them (which they never even acknowledged).
nVidia OTOH went out of their way to explain their extensions, listen to my suggestions, meet with me personally etc etc. "Outstanding support" would be a better description - they told me all I wanted to know (within reason), and I felt like they were listening carefully to the suggestions I had for future hardware. Haven't tried their ordinary tech support, if they actually have any (being a chip maker not a board maker like ATI).
...if you could LOCK the rendering at 30 FPS, and throw in motion and acceleration blur, it would still look better than a card rendering the exact same thing at 300 FPS.
No it wouldn't.
a) T-buffer's motionblur only did 4 subframe samples (V5 6k could do 8), which resulted in a very stepped-looking blur for even moderate motion, not a smooth blur at all. It'd look much the same as a card that rendered & displayed those 4 frames individually.
b) The card (and host CPU) still had to calculate, transform, upload, project & rasterise all those subframes, so the system had to be capable of 120 fps anyway. It's just as slow as a card that rendered & displayed those 4 frames individually.
c) A card that could render & display at 120 fps would show more detail than one that smeared the same thing down to 30 fps. It's much like anti-aliasing - great if you're limited by what you can display rather than what you can calculate, but not a substitute for real detail.
Motion blur isn't necessarily a good thing, it's just a way to convey movement information beyond the limitations of the display. The V5 was too limited in other ways (mostly speed) to take good advantage of this, but current cards could do a better job.
Since people's eyes/brains are generally limited to perceiving 60-90 fps anyway, there's a valid argument for using the ability to render at 300 fps to calculate subframes & create motionblur at 75 fps instead. Contrary to the ugly & overexaggerated blur that 3dfx liked to demo, this would actually help FPS players to track their target's movement at high speeds, by still conveying some of the extra information that a true 300 fps visual system could handle.
...you might have to modify...to get full support...
Clearly, you're no programmer. Imagining these things is a great deal easier, and requires much less time, than actually doing them.
I don't suppose you've wondered why it is that open-source programmers have been missing all these fabulous opportunities to cause MS pain? It's because the relatively few people with sufficient talent, time and inclination to do this sort of stuff are for the most part being paid to work full-time for a closed-source company instead...
Slashdot Mirror
Actually, 3DFX changed their own name to 3dfx, before the end...
Besides, although the FX Ultra may not sell well, the effort was definitely valuable. The Quadro FX (based on FX 5800 non-Ultra) is the easily the fastest professional card out there - it destroys the ATi FireGL X1 (9700 PRO-based) in speed, quality, precision and programmability (and yes, it's quiet). The standard 5800 also has a fair chance of success, if they can price it against the ATi 9500 PRO.
Also, nVidia have doubtless learned a lot about the 0.13um process, whereas ATi have yet to jump that hurdle (R350 will still be 0.15um). Rather than being behind, nVidia might just have a headstart...
I think that could be called "successful".
nVidia are a larger company with a string of huge successes to date. They have a much more diversified income, including some very popular OEM chips, the successful nForce2 (and less-successful Xbox) chipsets, a well-regarded pro card line, and a significant share of the Apple market too. Not to mention quite a bit of cash in the bank.
A single high-end chip(which is a small % of their total revenue anyway), even if it failed completely, is not going to impact their bottom line that much. It'll have more impact on their image as graphics leader, but they have the resources to learn, move on, redesign and try again.
It was out 3 days after the Xbox in the US, but 5-6 months *earlier* than the Xbox in Japan. That's where it got its lead from.
Completely false. The GameCube doesn't even have a digital audio out. The best it can do is ProLogic II. The PS/2 can do pre-encoded Dolby Digital 5.1, good for cutscenes and DVDs, but nothing in-game. A (very) few games do in-game encoding to DTS, but this requires dedicating one of the two vector processors to sound encoding.
With the Xbox, the Dolby 5.1 encoding is done by two dedicated DSPs on the sound chip, and is automatic for all games. It's the only one with the grunt (and the licence from Dolby) to do it.
The network adapter and built-in hard drive don't mean swat because nobody will be taking full advantage of them until the other two consoles get them.
There are ports of PS/2 games that have had custom soundtrack playback added, from the hard drive. There are a number of games that require the hard drive on the Xbox, Morrowind and Project Ego being a couple - PC ports may well require it. And of course, there are still Xbox exclusives that use it to good and occasionally innovative effect, such as Blinx.
XBox Live is making very good use of the ethernet port. So are tunneling apps like XBConnect that let you play peer-to-peer online multiplayer Halo/Mech Assault/Tony Hawk/Unreal/whatever. Can't do that with just an iLink connector.
Hell, most developers still make XBox games that look like ass and by the time they figure out how to take advantage of that hardware, the PS3 will be out.
And I suppose developers will instantly know how to use the (considerably more complex) PS3 hardware to best effect? Quite the opposite - look at how long it took before the PS2 hardware got used properly. Xbox has a big advantage here, with its far easier/more familiar dev environment.
Most PS2 games look like ass too. It depends on the quality of the game, of course. Halo is very nice looking, Rallisport Challenge looks superb, Splinter Cell is amazing and Halo2 is looking awesome. PS3 is still a while away, and Xbox2 is also scheduled for that timeframe.
Or this FotR parody.
Actually, nVidia have been doing this for a number of years too. Full stereo drivers (not some kind of hack) are available from their site that offer a lot of flexibility (e.g. different stereo methods, per-app convergence settings etc). The cards don't need a dedicated port, as one can simply use the vsync signal to flip sides - this works with pro equipment right down to $50 consumer LCD-shutter "toys".
Speed isn't an issue either - the cheapest GeForce MX card can do 1280x1024x32 at 150Hz (75Hz left & right signals) and can still pump out triangles & pixels a lot faster than the old Creator3D.
There were 3dfx-based consumer stereo setups in 1999, and I believe nVidia introduced their stereo drivers in 2000. Cards & cheap LCD-shutter glasses have been around for a long time; they were popular for a while as a novelty for playing 3D games.
Thing is, Sun equipment is expensive; it uses a number of dedicated ASICs and has a limited market, and is priced accordingly. The money isn't there for serious development. Consumer 3D cards generate far more revenue, and have therefore progressed amazingly in the last four years, to the point where ATI's and nVidia's consumer cards are now being sold (slightly tweaked) to professionals, easily exceeding all other professional cards in speed and features, without sacrificing quality.
It's great for 2D image processing. HDR images blur really nicely - the highlights stay bright, instead of going murkey grey.
Anyway, tiling as you describe is rarely used in motion picture image processing work, regardless of the number of layers. Breaking down a large (4000x3000 or larger) image does improve memory usage (sometimes at a cost in efficiency for certain algorithms), but when this is done, it's usually broken into scanlines or groups of scanlines, not square tiles. This works just as well and fits better with how images are processed, stored, displayed etc. The number of layers to be composited does not affect this at all.
DPX and Cineon do not support tiled image packing. TIFF does, but I've never seen a post-production app actually output a tiled image - it just complicates things unnecessarily.
And it's rarely necessary to re-read an entire image if you just want a subrectangle of it - many formats make it relatively easy to read a limited region. Compression can complicate things, but you can usually limit your reading to just the scanlines involved.
Most 3D packages render directly to high-dynamic range formats, including (I think) Povray. You don't need to render multiple exposures.
16 bit float is just one of the datatypes it supports. The particular format they chose is not limited to Cg or the GeForceFX, it's the most common 16 bit float format out there, even if it isn't an IEEE standard.
The DirectX HLSL is (deliberately) syntacticly identical to Cg, so that's actually good for Cg, rather than killing it off as you suggest. OpenGL2's HLSL has yet to be confirmed, but if (as may be likely) it isn't also just like Cg, Cg will still be able to compile to OpenGL2 - it's just another render target, along with DX8/9, OpenGL 1.3/1.4, and nVidia's own extensions.
HDR info is useful for many many things in both 3D and 2D work (though I'm doubtful about edge detection). Other HDR-supporting formats do exist, including HDRI, TIFF, FLX, and RLA. Even Cineon/DPX supports limited HDR info. Each have their own advantages & disadvantages - OpenEXR is no worse than most, and better than many.
Rendering technique is only one small part of the whole job. If you want to take HDR info from one device/app/system to another, you have to write it into a file, so you need a file format that won't clip all your highlights...
It was nVidia's move to 0.13 micron that delayed the GeForceFX, and allowed ATI their moment in the sun. ATI have yet to climb that particular hill, and nVidia are already rolling down the far side.
Is there something magic about the Creator3D cards that I don't know about? According to Sun, they could do 3D at 1280x1024, had 15MB of framebuffer RAM, and could only do 1.4M tri/sec. There's no indication that it even supported texture mapping.
At first glance, the consumer market had already caught up in 1999 - the GeForceDDR comfortably exceeded those specs. These days, an $80 gfx in your Dad's new PC could probably beat it out. Heck, even the integrated gfx in an nForce2 board has higher specs.
There's nothing specific to Windows that prevents multiple capture cards. I can choose to capture from my ATI AIW's S-Video input, composite input, TV tuner or Firewire port, or my nForce2 mobo's Firewire port. The new AIW also allows plugging in an additional PCI capture card (TV Wonder) and getting PiP functionality, or watch one/record another.
The service has parental lockout, so you can give your credit card and not worry about your kids racking up huge bills for premium content.
Clarification: AMD asks us to point out that Hammer schedules haven't slipped from its previous advice, as we originally suggested in this article. A spokesman from the company told us that desktop versions of Hammer are still planned to ship (for revenue) in Q1 2003 with systems on shelves at the turn of Q1 2003, not the second half of 2003 as we stated.
They're saying that Barton will be here 1Q03, Sledgehammer is due 1H03, but now ClawHammer may be delayed until 2H03!
Arghh. I thought the point was to do a 64 bit CPU without requiring an Itanium schedule...
While an 11 Mpixel sensor does have 11 million spatially separate elements, they are monochrome, filtered to capture R-G-B or C-M-Y, and cross-interpolated to provide a true colour image. This gets you close to 11 Mpixels of luminance resolution, but almost a third of that in chroma resolution. A RAW format image from such a sensor could probably be exported to a pretty sharp B&W image though, good enough to resolve alternating black & white lines.
There's also the Sigma SD9, using a Foveon X3 sensor. Only 3.4 Mpixels, but with RGB elements for each one, stacked vertically. That means the chroma resolution is almost that of the Canon 1Ds, but without any artifacting caused by the cross-interpolation of the colour elements. It more closely mimics film. Can't wait to see what a 14 Mpixel Foveon sensor can do :-)
b) You won't get anything usable from scanning 35 mm film beyond around 4000 x 3000 anyway, with most film stocks - the grain overwhelms the pixel size.
c) The Canon 1Ds (and Kodak 14n) have 12 bit sensors, which gives a dynamic range of 1:4096.
d) The Kodak DCS 14n is built with a standard Nikon SLR lens mount. The Canon EOS 1Ds is compatible with over 60 of Canon's EF lenses.
OK, a decent SLR is a lot cheaper, but it doesn't have any of the advantages a digital camera gives you, either.
That indicates that the film was shot anamorphic and cropped for the laserdisc to be 4:3.
With ATI, they answered my questions accurately but briefly, and made no particular response to the bug reports I posted (which finally seem to be fixed now - mostly - in 02.3, 8 months after I notified them). Their dev support team gave me adequate support, IMHO. Their regular user technical support was not even close to adequate, the once or twice I've used them - I had to track down the fix myself and explain it to them (which they never even acknowledged).
nVidia OTOH went out of their way to explain their extensions, listen to my suggestions, meet with me personally etc etc. "Outstanding support" would be a better description - they told me all I wanted to know (within reason), and I felt like they were listening carefully to the suggestions I had for future hardware. Haven't tried their ordinary tech support, if they actually have any (being a chip maker not a board maker like ATI).
True, your Xbox has to be modded to run these, and they're harder to find, but they've been around for some time now.
No it wouldn't.
a) T-buffer's motionblur only did 4 subframe samples (V5 6k could do 8), which resulted in a very stepped-looking blur for even moderate motion, not a smooth blur at all. It'd look much the same as a card that rendered & displayed those 4 frames individually.
b) The card (and host CPU) still had to calculate, transform, upload, project & rasterise all those subframes, so the system had to be capable of 120 fps anyway. It's just as slow as a card that rendered & displayed those 4 frames individually.
c) A card that could render & display at 120 fps would show more detail than one that smeared the same thing down to 30 fps. It's much like anti-aliasing - great if you're limited by what you can display rather than what you can calculate, but not a substitute for real detail.
Motion blur isn't necessarily a good thing, it's just a way to convey movement information beyond the limitations of the display. The V5 was too limited in other ways (mostly speed) to take good advantage of this, but current cards could do a better job.
Since people's eyes/brains are generally limited to perceiving 60-90 fps anyway, there's a valid argument for using the ability to render at 300 fps to calculate subframes & create motionblur at 75 fps instead. Contrary to the ugly & overexaggerated blur that 3dfx liked to demo, this would actually help FPS players to track their target's movement at high speeds, by still conveying some of the extra information that a true 300 fps visual system could handle.
Clearly, you're no programmer. Imagining these things is a great deal easier, and requires much less time, than actually doing them.
I don't suppose you've wondered why it is that open-source programmers have been missing all these fabulous opportunities to cause MS pain? It's because the relatively few people with sufficient talent, time and inclination to do this sort of stuff are for the most part being paid to work full-time for a closed-source company instead...