Slashdot Mirror

I don't pay close attention to the GPU market in general, though lately I've been interested in a few numerical modeling projects that could benefit from high-performance computing. The AMD Firestream 9170 is supposed to be released in the first quarter of this year, with a peak speed of 500 GFLOPS, most likely single-precision, but the beauty part is that it should also support double-precision, the numeric standard for most computational modeling. NVidia's option in this space is the Tesla C870; I wonder whether this move to purchase another GPU line will divert resources away from their number-crunching-first GPUs.

I don't believe AMD/ATI. their video cards have always had really bad unix support.

Quite true. Many years ago, almost last century ATI had a series of well supported cards that worked well with Linux, but also BSD and Solaris. Quite nice cards too in their time. But then ATI changed hard and fast to being closed and getting drivers became near impossible. A lot of times the VGA basics would work, but you were under utilizing the card. It is about where I stopped buying ATI and moved to nVidia because they are least supply drivers. nVidia drivers here Given ATI support is increasing, I will watch. But just that.

http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-13.html
http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-14.html
http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-15.html

These three sections will give you everything you should need to configure that FX5200. Note the selection on the bottom of each chapter has a link to the index... lots of good info I would suggest reviewing regardless of your usage of those features.

http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-13.html
http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-14.html
http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-15.html

These three sections will give you everything you should need to configure that FX5200. Note the selection on the bottom of each chapter has a link to the index... lots of good info I would suggest reviewing regardless of your usage of those features.

http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-13.html
http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-14.html
http://us.download.nvidia.com/XFree86/Linux-x86_64/169.07/README/chapter-15.html

These three sections will give you everything you should need to configure that FX5200. Note the selection on the bottom of each chapter has a link to the index... lots of good info I would suggest reviewing regardless of your usage of those features.

Blue/Gene L is rated at 500 TFLOPS, which is impressive, however if you don't need double-precision to do this stuff, you can run very fast on much cheaper hardware. I was looking at Nvidia Tesla cards and boxes recently, and those are claimed to pump out 500 GFLOPS per CPU... with a 4 CPU device (1 TFLOP) taking up 1U of rackspace. I think this technology will ramp up a lot faster than people expect.

Crap article - and it sounds as if this fellow should put in a resume to Evolved Machines - http://www.evolvedmachines.com/

An Nvidia spot....*drum roll*....featuring neural simulations on GPUs
http://www.nvidia.com/object/tesla_testimonials.html

No, I mean a RAID controller built onto the motherboard using shared resources on the southbridge. http://www.nvidia.com/object/feature_raid.html

PowerMizer Mobile Technology page: http://www.nvidia.com/object/feature_powermizer.html

Maybe the NVIDIA Technical Brief will yield some answers: http://www.nvidia.com/object/IO_26269.html (Warning, spawns a PDF)

PowerMizer7.0 Power Management Techniques:
Use of leading edge chip process
CPU load balancing
Intelligent GPU utilization management
Revolutionary performance-per-watt design
PCI Express power management
Aggressive clock scaling
Dedicated power management circuits
Display brightness management
Adaptive performance algorithms

CPU Offload Example (from NVIDIA's Technical Brief)
Figures 3 and 4 (see PDF) show CPU utilization when running a Blu-ray H.264 HD movie using the CPU and GPU, respectively. You can see that under the GPU video playback, 30% less CPU cycles are being used. This dramatic reduction in CPU usage means less power is being consumed by the processor, therefore system power consumption is reduced. resulting in longer battery life.
Note: Testing was conducted on an Intel Centrino based platform with 2 GHz Core2 Duo processor, and a GeForce 8600M GS, running Intervideo WinDVD8 playing a Casino Royale H.264 Blu-ray disc.

PowerMizer Mobile Technology page: http://www.nvidia.com/object/feature_powermizer.html

Maybe the NVIDIA Technical Brief will yield some answers: http://www.nvidia.com/object/IO_26269.html (Warning, spawns a PDF)

PowerMizer7.0 Power Management Techniques:
Use of leading edge chip process
CPU load balancing
Intelligent GPU utilization management
Revolutionary performance-per-watt design
PCI Express power management
Aggressive clock scaling
Dedicated power management circuits
Display brightness management
Adaptive performance algorithms

CPU Offload Example (from NVIDIA's Technical Brief)
Figures 3 and 4 (see PDF) show CPU utilization when running a Blu-ray H.264 HD movie using the CPU and GPU, respectively. You can see that under the GPU video playback, 30% less CPU cycles are being used. This dramatic reduction in CPU usage means less power is being consumed by the processor, therefore system power consumption is reduced. resulting in longer battery life.
Note: Testing was conducted on an Intel Centrino based platform with 2 GHz Core2 Duo processor, and a GeForce 8600M GS, running Intervideo WinDVD8 playing a Casino Royale H.264 Blu-ray disc.

It appears Alienware will be using the GeForce 8800M GTX in their "m15x" and "m17x" models:
http://www.alienware.com/intro_pages/m17x_m15x.aspx
NVIDIA GeForce 8800M Link: http://www.nvidia.com/object/geforce_8M.html

Nothing like primary source material, folks.

I really grow tired of Slashdot linking to another site that describes a web page, instead of just linking to the page itself.

There are other recent issues with nVidia that are causing people to consider "jumping ship" and going with ATI for their next card.

nVidia has PureVideo, but afaict you have to buy the software from nVidia or from a software partner. OTOH ATI has native support included with their drivers (named AVIVO).

I believe ATI's support for linux is also moving ahead, including support for hardware decoding under linux. However with nVidia you have fat chance of using "PureVideo" under linux.

Although this is possibly irrelevant considering most people that buy this card probably don't even know that you can run games in linux, but the increasing spread of High Definition videos will make these issues much more prominent.

Sorry for replying to my own post, but some more digging turned up this page with details on their HPC products:
http://www.nvidia.com/object/tesla_computing_solutions.html
and this is the card with 128 parallel processors and 1.5GB memory:
http://www.nvidia.com/object/tesla_gpu_processor.html

Sorry for replying to my own post, but some more digging turned up this page with details on their HPC products:
http://www.nvidia.com/object/tesla_computing_solutions.html
and this is the card with 128 parallel processors and 1.5GB memory:
http://www.nvidia.com/object/tesla_gpu_processor.html

Video cards don't have 128 processors exactly, but each processor can have 128 separate instruction pipelines (SIMD 'stream' processors) which can run in parallel http://www.nvidia.com/page/geforce_8800.html. The programmer just loads a small amount of code into each of these stream processors and lets it crank away at the data. In graphics processing, these stream processors are used for per-pixel and per-vertex shading.

Why do NVidia cards that can run CAD/OpenGL cost several times as much as NVidia cards that can't?

I have no idea what you mean. Take the Quadro FX 3500 - nVidia classes it as 'Ultra high-end' on their workstation cards, and compare it to a GeForce 8800 Ultra, which is marketed typically to gamers:
Quadro / GeForce
memory: 512M / 768M
memory interface: 256bit / 384bit
memory bandwidth: 42.2 GB/sec / 103.7 GB/sec
BOTH cards support OpenGL 2.0 and DirectX 9.0c.

So there you have it, a high end gamer card that performs on par/slightly better than an 'ultra high end' workstation card. Priced about the same at reputable retailers. I can not find an instance where a comparable "CAD/OpenGL Card" costs "several times" as much. Mind sharing? Or are you complaining that CAD is more graphics intensive than video games? If so, it wouldn't matter what API you use... DirectX, OpenGL. Just a programming API to the same damn video card.

Why do NVidia cards that can run CAD/OpenGL cost several times as much as NVidia cards that can't?

I have no idea what you mean. Take the Quadro FX 3500 - nVidia classes it as 'Ultra high-end' on their workstation cards, and compare it to a GeForce 8800 Ultra, which is marketed typically to gamers:
Quadro / GeForce
memory: 512M / 768M
memory interface: 256bit / 384bit
memory bandwidth: 42.2 GB/sec / 103.7 GB/sec
BOTH cards support OpenGL 2.0 and DirectX 9.0c.

So there you have it, a high end gamer card that performs on par/slightly better than an 'ultra high end' workstation card. Priced about the same at reputable retailers. I can not find an instance where a comparable "CAD/OpenGL Card" costs "several times" as much. Mind sharing? Or are you complaining that CAD is more graphics intensive than video games? If so, it wouldn't matter what API you use... DirectX, OpenGL. Just a programming API to the same damn video card.

Depending on your application, you don't need low-level access to the GPU to utilize shaders for computation. Check out some of Nvidia's tech demos for their CUDA stuff - not saying that the PS3's GPUs support that kind of stuff, but if your software can read shader output, you can utilize basic shader functionality to perform certain kinds of transformations very quickly (cross/dot product, matrix multiplication, etc).

A couple of the CUDA examples include a Mersenne twister implemented on the GPU and fluid dynamics simulations.

CUDA aside, it's been possible for quite some time to do things like edge detection, producing histograms and frequency analysis on images using only programmable shaders. Additionally, the GPU is fairly well suited to preparing smooth surfaces from MRI images in realtime.

There are official forums, but it seems like NVIDIA doesn't care. It is REALLY frustrating. I think I will be going back to ATI/AMD for my next video card.

It's very cool that the memory becomes available so easily with just a couple driver parameters. It's a pity that there's a lot to optimize before it can really shine.

Memory architecture on a GPU is very different from system memory. Memory there is not linear and the video memory controller will go through a lot of remapping to present it as such, something that's probably very slow because of the VBIOS. Then there's the issue of tuning the bus so that reads and writes are using its full bandwidth, and again a poor VBIOS implementation may be the bottleneck.

The best but harder solution would be to have a means to program the video memory controller directly to map pages of system memory and do all the copying and moving itself. Of course, this is hardly ever going to happen, but some improvements can still make it into the VBIOS, some of which will probably happen once GPGPU-style programming starts getting more attention as both nVidia and AMD/ATI are seemingly interested in pushing with things like CUDA and Stream Computing.

The concept as it is now, however, remains extremely cool. It might still be orders of magnitude slower in terms of latency and throughput compared to system memory, but it should be a lot more responsive than a hard drive just because there are no seek times involved. That said, hdparm -t may not be the best tool for measuring performance, so i'd be more interested in a random access benchmark since it may make some use of the parallel memory architecture inherent on a video card.

I am curious anyone has ported gnugo (or whatever) to the nvidia cards using their CUDA C Compiler. The high end nvidia cards are something like 128 500 mhz processors. If a position was "at clock" you would only need 20 nvidia cards (and some coding to glue them together) to get 10^12 positions / sec. If it takes say 50 instructions to evaluate a position, that means 100 nvidia cards gets what you need or 20 boxes w/ 5 cards in each, it may not be "at home" affordable but its certainly on the low end w/ regards to cluster style computing. The coding would be a pain but I would guess that the tools are there (MPI wrapped cuda?)

Since day 1. This is the point of non free drivers. Some makers are better than others at "supporting" new and "legacy" Windoze, but you will never get everything you need unless you buy everything new every three years and throw away the old. It's intentional waste and that's what non free software is all about.

With raytracing, there are lots of new possibilities. For one thing, reflection and refraction actually work like they do in real life.

Single-bounce reflections and refractions can be done without raytracing. Now, reflection through a glass sphere is more interesting than a single bounce (and embarassingly easy with raytracing), but if you look at NVidia SDK Samples, you'll see plenty of non-raytraced examples that do use single-pass traces (in a pixel shader, usually).

Lights can work accurately if you want them to

Area lights? Or point lights? Light maps are (in general) a good and simple solution to dynamic lighting, but they suffer from pixelation problems in some situations. Thus, care is required to prevent lights from getting into "bad situations", but the rendering is still faster using cache-friendly rasterizating. Dynamic area lights are more difficult for raytracing and rasterizing alike, so the general trick is to use static (precomputed) area lights and dynamic point lights.

...and radiosity can be precomputed for static scenes.

This can be done regardless of the rendering method. Now, realtime radiosity, that would be sweet, because it can be combined with either raytracing or rasterizing. I don't have a link handy, but Henrik Jenssen (sp?) has a GPU-based photon-mapping solution, which would add caustics. I doubt that would scale up to an appropriate quality for gaming right now, but sooner or later someone will figure out a hack that's close enough.

Most of it (except good lenses) has been faked before with rasterization, but raytracing will actually let you set up a series of mirrors and telescopes to peek around corners in a FPS for instance. I can imagine a true hall of mirrors in an FPS would be at least a little more interesting than what we have now, too.

That certainly would be cool. I betcha a hybrid solution would be the fastest, though. Raytrace just the dynamic objects in a scene through the mirrors, as a CPU-based vertex processor... but still render by rasterizing. After all, dynamics systems for collisions are already similar to raytracing... aside from having a physics subsystem, maybe some games could start implementiong a raytracing subsystem prior to rendering.

Raytracing is O(log n) versus O(n) for rasterization, which means that even though raytracing is currently slower (the constants involved in raytracing are higher), after the break even point is passed much less of the available computational power will be needed to render the scene and can instead be used for physics and AI.

The thing is, even as raytracing gets faster, rasterizing gets faster, and people are still pushing new research into rasterizing, and GPUs are continually speeding up. Rasterizing algorithms are much more cache-friendly, which is one of the biggest weaknesses of raytracing that will never go away. GPUs presently blow CPUs out of the water, performance-wise, so you're at the point of waiting for N-way processor machines to be on desktop machines. NVidia and ATI aren't going to sit around with their thumbs up their asses while they wait for CPUs to put them out of business, which means anything that can run on a GPU is still going to outpace the multicore CPU systems.

No matter how fast someone makes a raytracing solution, someone else is going to jump through whatever hoops are necessary to have a faster rasterizing solution, given the current trends in hardware. I suppose this all goes out the window if a GPU is developed that is better at raytracing than rasterizing, though.

...and I have two Nvidia 8800's running in SLI which also is windows only
'Windows' doesn't support Nvidia hardware, Nvidia does, with excellent drivers for Linux and XP (and allegedly less than excellent drivers for Vista...)

http://us.download.nvidia.com/XFree86/Linux-x86/100.14.11/README/chapter-25.html

I also have an Geforce 4. Its a Geforce 4 MX 420, a very old model. But its still working and I'm happy with it. There's only one bug working with Beryl. Nvidia told me that they are on the way for a solution. Why I'm responsing is to inform you that it is possible, even with an old Geforce 4 to get the new kernel. All you have to do is to install the Nvidia Legacy Driver. They are still supported and updated, just not as frequently as the normal driver. You can get the legacy driver here: http://www.nvidia.com/object/linux_display_ia32_1.0-9639.html Good luck with them!

And the newer drivers still support chips that predate your GF4.

For the record, Nvidia says otherwise.

"Thank you, oh benevolent masters, for supplying the software required to use the hardware that you gave me in exchange for money." Was that suitably deferential, or should I bend my knee more?

As a pleasant side effect for my /.'ers, desktop support will rock too. But I don't think that's what this is about.

It's not google's job to fill in the gaps. The distro will take care of that. Always has for the thousands of applications already out there for every linux distro.

The beauty of the dynamic linker is that you can have multiple library versions co-exist quite peacefully together. If OpenGL games can run on multiple linux base distros, why can't (of all things) Picassa? Or [insert gApp here]? Worst case, google can link to RPM(s), DEB(s), or whatever they already pre-packaged themself. Binary globs? Do I really need to illustrate the process? Shoot, closed source binaries? I think google can handle it. I think your real argument lies with BSD, not linux.

I was pointing to the Nvidia deskside tesla supercomputer they sell here http://www.nvidia.com/object/tesla_deskside.html I made no reference to the 8800GTX. And it's good to see some people actually due this kind of work. Last time I did anything with iterative was BAM neural networks and that was 15 years ago. As you said it, a GPU based supercomputer would be great at non-conditional logic and a boon to AI research.

I thought the hip thing was GPU based supercomputing. NVidia even has a dedicated GPU based, desktop sized, scalable supercomputer line called Tesla.

The basic Tesla unit c870 = 518 Giga flops for ~$1300.
Tesla s870 = 2 Terra flop for ~$12000 (still desktop size)

NVidia Tesla

Although not as cheap as the Microwulf, Nvidia has a desktop super-computer for sale http://www.nvidia.com/object/tesla_deskside.html at 500 GigaGlops, to start.

> If you look at nVidia cards, only the 5 (FX) series and later support GL 2.0, the earlier ones are 1.5 only. Why? They can't accelerate GL 2.0. Rather than have it implemented in either a semi-working fashion, or a slow software emulation, you just support the maximum level you can.

That's not really true.

Actually, you can download the Windows "nvemulate" utility for free and request your NVIDIA OpenGL driver emulate a level of GPU functionality beyond what your actual GPU hardware supports. You pick what GPU generation you want and you get all the OpenGL extensions for the GPU you are emulating. That includes getting OpenGL 2.0/2.1. All rendering that can run fully hardware accelerated continues to do so. See:

    http://developer.nvidia.com/object/nvemulate.html

A fair number of students and developer using their laptops use this functionality to develop code for GPUs beyond what they have.

Caveats:

1) There's no support guarantee from NVIDIA for this mode.

2) It's pretty darn slooow if you fall back to software for most features.

3) nvemulate only applies to OpenGL, not Direct3D.

- Mark Kilgard, NVIDIA

However, to play a 720p file that is encoded with h.264, you need some serious punch in the processing realm.

Can't you leave that to the video card?

Very nice game indeed, but I am not going to get the full game...

I also got to play it tonight (I actually had free time and not in crunch mode?) for about 40 minutes (yes, it is short). The previews, screen shots, video clips, and trailers didn't excite me for this game. I kept hearing and reading very high scores from Xbox 360 port (demo and the full game that was sold earlier). Everyone was raving how scary, addicting, and pretty the game was. Now, I know why. The audio, graphic, special effects, etc. were very nice.

POSSIBLE SPOILERS: This 3D surrealistic first perspective shooter (FPS) game and story theme was an issue for me since it didn't hit me to excite me. It takes place in 1960 in an underwater city (it reminds me of Atlantis, Titanic, Blade Runner movie, etc.). The demo started out with an introduction that reminded me of Lost's Oceanic Flight 815 jetliner crash in the sea/ocean, but at night time. Wow, looking at the water was LOVELY and seeing the water splashes and droplets on my screen! While swimming to the lighthouse near by, I heard the flames, explosions, me coughing out water and breathing, etc.

The fun start begins in the lighthouse when I travel down to the underwater city named Rupture. At the same time, I met a guy helping me over the radio. You can hack robots to be on your side and protect you, security cameras, sentry guns, etc. There are various life spawn spots if you die. If you played System Shock 2, then you would recognize that this is the same people who worked on this game. The whole game system is based on it, but on a different game engine. The game still had scary parts, beautiful graphics and effects, objectives/missions, etc. It also reminded me of American McGee's Alice 3D FPS game for the surrealism and weirdness.

Check out the game if you have a decent gaming system or a Xbox 360 (heard it was good on the console as well and there's a free demo). Enjoy the graphics, special effects (check out those neat water falls, leaks, etc.), cutscenes, sounds, music, and horror. I was surprised it ran well on my not super fast system even without the beta NVIDIA driver that is supposed to be supported for this game.

Circuit City weekly ad/advertisement shows it for $39.99 for this week. So one extra copy for you to buy since I am not buying it due to above reasons and lack of free time (got other games to play and finish). If it was a sequel to System Shock 2 game, then I would be all over it just for SHODAN (I miss her harassing me like saying "Look at you, hacker: a pathetic creature of meat and bone, panting and sweating as you run through my corridors. How can you challenge a perfect, immortal machine?")!

> and major requirement change

It is not a major requirement change, because, contrary to the statement of The Inquirer, the previously optional and now mandatoryfeatures are provided by NVidia (source) and ATI DX10-cards (source).
Both are have 32-bit fp unified shaders and 4xAA.

Looks pretty compelling from here! Take a look at the 32-bit vs. 64-bit benchmarks shown here for instance! If I was a game developer, I'd gladly take 40FPS over 30FPS if it only meant a recompile targeting a 64-bit platform!

That's nowhere near representative

E.g.

http://forums.nvidia.com/index.php?showtopic=29278 &mode=threaded
I get unstable results in Vista x64 the range from as low as 4800 to as high as 5200, this is a massive instability when you consider that ins Vista x32 I get 5270 +/- 20 and on XP (both x86 and x64) I get 5300 +/- 15. If anyone would like the compare URLs I can post them, just ask

x64 seems to have a graphics driver issue.

or

http://news.softpedia.com/news/Windows-versus-Wind ows-or-32-bit-versus-64bit-1349.shtml

x64 and x32 perform identically. This is what I'd expect actually.

You only need to test two binaries if you also choose to support 32-bit as well - a suitably advanced app/game could just make a 64-bit capable AMD/Intel chip a prerequisite these days (DOOM required a 386 or better during a similiar 16/32-bit transition period)

Yeah but back then every gamer had a 386. Now everyone has an x64 compatible CPU but most of them are running 32 bit Windows.

PAE only gives you access to more memory, it doesn't enable the CPU 64-bit processing

x64 gives more registers, and you can use things like CMOV since you know all x64 CPUs support them.

Having more registers won't help that much I suspect - most games are limited by GPU or memory subsystem performance, not by CPU registers.

I see absolutely no need to continue developing the desktop (workstations are another story). Here is what is out for the laptop.Core Extreme from Intel. 250 GB hard drives from various vendors. Great graphics solution from Nvidia and AMD. Notebooks can be fitted with up to 4 GB of ram. Notebooks utilized power more efficiently which is important in this energy conscience age. I personally have no need for desktop and will never buy one again. A laptop combined with a good display is enough. It is time for the OEM to get clever and start experimenting on new designs built around mobile components. Imac is a good example of a destop built from laptop components. Maybe special docking station for the hardcore users who want more expansion that the express card slot. A eSATA for notebooks perhaps for extra hard drives.

Lies, lies and more lies. How is Microsoft screwing Apple in this situation?

The iPhone doesn't have to interact on any level with DRM, and the component parts of my system that do interact with it (audio, video) have full 64-bit driver support. I should know, I'm using 64-bit Vista. Even better, XP doesn't have any of the same support for Blu-ray/HD-DVD DRM, so how are they not able to code XP 64-bit drivers? Apple have no excuse in this situation when plenty of hardware manufacturers are able to code them.

Then you reel off the same stupid list that you repeatedly cite to show that 'M$ am bad' which has been debunked a hundred times by Windows and Linux users alike.

Your final sentence is a gem that sums up the rest your post: "Windoze is like barren". What does that even mean?

This week was IVR, the industrial virtual reality show, in Tokyo. SGI was showing a 4K SXRD projector (FWIW a 35mm slide is 8K, if you printed one at 4K it would suck, but 4K is 4 times better than high def.) and despite their very annoying switching between showing fact sheets in 2D (so your right eye felt blind in the polarized glasses) and 3D, what was cool was the window into the room to see what was driving it. They were using their Asterism high-end Windows machines (I think maybe 4, at least that was what the picture showed) networked to some number of NVIDIA Quadro PLEX rendering units (that's what I want for Christmas!). Having seen their infinite reality systems in the past I was not totally stunned, and in fact they had a very annoying ferrari 3D model that stuck so close to your eyes it hurt, but clearly they had built an interesting rendering network.

Many other exhibitors were also using these NVIDIA standalone rendering bricks and one booth I noticed was using two Dell PCs to drive a stereo view. Some were using 2 projectors superimposed on each other for higher contrast. So this stuff is all here already, heck I saw a Hi-Vision system using two superimposed projectors like 10 years ago, but easy self configuring and aligning systems that would let you plug in more pcs, rendering resources or projectors would be very cool.

Many other

..So what if it were to be outfitted with 72 racks full of these?
http://www.nvidia.com/object/tesla_gpu_server.html
"...The Tesla S870 GPU computing server is used in tandem with multi-core CPU systems to create a flexible computing solution that fits seamlessly into your existing IT infrastructure."

Surely 294,912 processors are not too many for this, but wait!, CUDA is programmed just like C... No problem!

"Buying an $700 DX10 Video card before there are any DX10 games means that you will pay a lot for a card that will cost a lot with very little return."

CUDA

I haven't heard that analogy before, and I don't believe it's correct (though if you mean creating unique trees after placement, possibly yes). It's a bit more primitive than plopping the same model down in a different location.

First off, there was a terminology issue for a while where OpenGL used primitive shader instead of geometry shader, but this seems to have been resolved (from a search, the formal extension is GL_EXT_geometry_shader4, which usually is kept in the ARB/core name), so you may find information under either primitive shader or geometry shader for OpenGL - just a heads up. The OpenGL 2.1 Specification, which is current on most cards that include DirectX 10 DOES NOT have to include geometry shaders since they are still considered extensions (EXT) and not ARB or core (which are required); however nVidia does include the extension in their 8xxx cards and I suspect ATI does in their DX10 cards as well. For more formal information, you may want to search the Khronos site for Long's Peak (which will be 2.2 or 3.0) or Mt Evans (likely 3.0 3.1) releases.

Geometry Shaders are essentially a higher level object than vertex shaders (executed after the vertex shader on a set of vertices and before fragment processing). For example, if you take a triangle strip (the highest level object considered a "primitive," and probably the reason they're called geometry shaders instead of primitive shaders now) forming a cylinder and want a flare at the bottom, you might add vertices to keep it fairly round and move the existing vertices along the bottom ring outward. That probably isn't the main use for it, but is an example (other uses: cube mapping and probably the #1 I've heard of, CPU independent particle systems not limited to sprites)

In 20 years, we'll smell our avatar's farts. Huge advance. As a result, we'll all be having sex in 3d virtual worlds instead of the real world. For the first time in human history, recreation will be fully separated from procreation.

http://www.buzzle.com/articles/day-smelly-computer -has-arrived.html/ + http://www.fuckingmachines.com/ + http://www.nvidia.com/page/home.html/ = http://www.3d-sexgames.com/

Probably won't take 20 years. All the pieces are there, just waiting for some shameless pervert to assemble and market them.

Being fabless works for lots of companies, for example NVIDIA (disclaimer: I work for the gentle green giant).

There are lots of companies who only do fabrication, just as there are many other fabless semiconductor companies. With process shrinks occuring as quickly as they are today, it makes a lot of sense to let someone else (or several other someone elses) deal with the cost of developing fab facilities capable of the latest and greatest process size.

Note, I'm not a lawyer. Contact one today.

1. Can I do it with Linux today (GPL2) and tomorrow (GPL3)?

YES. Nowhere in the GPL, ether version 2 nor 3, precludes you from running a closed source binary on a Linux system. If it did, we wouldn't be as far as we are now -- and I wouldn't be running Second Life on a Nvidia Geforce 7 series card.

2. Can I statically link the code with Linux libraries? (My own experience shows that dynamic linking is too much to bear.)

Yes, check the manual pages and info pages on GCC. To keep on the licence issue, the GPL does not transfer to code compiled by GCC, and the "Lesser GPL" that is used on libraries allows you to staticly link them into closed code.

3. Can I obfuscate my code (e.g. encode it)?

It would be useless. If you ship a binary only, the obfuscation would already be worked out by the compiler, and a decompiler could be used to steal the code. Since you're not licencing under the GPL, you must make a licence that prohibits decompilation and disassembly without permission.

4. Could I be forced to publish this code by some 3-d party?

No. Since you are not licensing it under the GPL, the BSD variant licenses, the Apache license, the Perl Artistic License, or any Open Source Foundation approved license, it would be under what you specify in your own license.

5. Am I correct that programming in and selling BSD-based boxes won't raise any of the above problems?

It is only a legal issue, not a technical issue. You really need to consult a lawyer who is willing to delve into the GPL, the Lesser GPL, and the BSD licenses.

Where is nVidia's bug reporting mechanism BTW? I've never been able to find one. Their website says if you purchased your product directly from nVidia, fine, but otherwise talk to your OEM. They used to have a feedback form page that was always 'under construction'. They've removed that entirely now (although just noticed added a new one for developers): http://www.nvidia.com/page/support.html

The infinite loop bug is hellacious, but either you've got it or you don't.

> As for Intel, my desktop at work has an integrated Intel chip which doesn't seem to work properly with Xorg 7.2
Intel's integrated i9xx chipset has 3D built-in. I've got it on my lappie. It's not a bad implementation, but definitely lowest common denominator and won't play newer games. When they announced their re-entry to the 3D field, they got a lot of flack for this chipset. I'm hoping Intel's 3D team will feel the heat from that. (I'm naive, but occasionally good things happen . :-)