Naked Science Forum
Non Life Sciences => Geek Speak => Topic started by: CliffordK on 13/02/2012 10:15:34
-
One of the things that has been ballooning over the last few years is the Video RAM (VRAM). It hasn't been too long ago that a couple of MB might have been common.
Then a few manufactures came up with the bright idea of shared memory.... which actually seemed to be quite SLOW!
Anyway, I'm now seeing cards from 512MB to 2GB cards for quite cheap. And, if one really wants to go whole hog, one can get 4, or 6GB VRAM. Double that if one puts in multiple cards.
But is all that really necessary?
If one takes a monitor 2560x1600x4 bytes, that comes up to 16,384,000 bytes. or just under 16 MB. So, with 2GB, I should be able to represent the entire bitmap 131 times.
I can dynamically monitor my main CPU load, as well as main memory usage, but I don't think there are built-in GPU Monitoring utilities.
Ahh, found a Windows nVidia&ATI GPU monitor... (http://www.windows7download.com/win7-gpu-monitor/mokttrar.html)
As well as notes on checking the nVidia GPU status in Linux (http://forums.nvidia.com/index.php?showtopic=63144). We'll see once my new card shows up.
-
It has more to do with image processing than storing 131 representations of your monitor, though how much you need depends on what you use your computer for. Most modern 3D games have to do a lot of processing to render their graphics, which is what has driven this boost in VRAM. It's a matter of computing where each polygon moves in 3D space from frame to frame, figuring out what 2D scene to render from that 3D information, and then applying all sorts of filters and visual effects to it, and then doing that 20+ times per second.
If you do high-end graphics processing for a living, you also find that these programs take huge amounts of RAM to store high-res images which may have more pixels than your desktop monitor.
Finally, a nice side benefit of the gaming craze is that these graphics processors can also be appropriated for a lot of scientific computation. If you can split up whatever computation you need to do in the right way, a $200 graphics card can be nearly as effective a high end workstation, since it has a lot of processing power and a lot of RAM and is designed to do simple computations extremely quickly.
-
... a nice side benefit of the gaming craze is that these graphics processors can also be appropriated for a lot of scientific computation.
or cracking encrypted wi-fi [:0] ... http://www.tomshardware.com/news/elcomsoft-gpgpu-wifi-crack,6863.html
-
Interesting...
I see I need to figure out how to hack a GPU.
The fastest Intel Pentium derivatives are running about 50 to 100 gigaflops. (billion floating operations per second)
The nVidia Quadro 6000 is running 1000 gigaflops (single precision), or 500 gigaflops (double precision).
And, 144 GB/s memory bandwidth for the nVidia, vs 20 to 30 GB/s for the Intel CPUs.
So, this would be essentially RISC Computing. (http://en.wikipedia.org/wiki/Reduced_instruction_set_computing)
Probably some form of a parallel RISC.
Hmm, the nVidia GTX 580. memory bandwidth 192.4 GB/s, and 1581 GFLOPS. Whew!!! And the GTX590 is essentially a dual processor form of the 580.
-
The NVIDIA GPU programming platform is called CUDA, so you might want to start there. It's the one I've heard used the most in scientific computing. I've seen some demos of what it can do, and its pretty amazing: I've seen computations that might take a minute or so done in real time at maybe 10 FPS.
-
Wasn't the fastest supercomputer in the world (for a while) a helluvalotta playstation cell processors all strung together?
-
Yeah, it was a US military project, I believe. Video gaming is a huge industry now, so it drives computing power faster than science. Now scientists find themselves formulating problems in a way that video game technology can be applied to them! :p
-
So, does this mean we are really all just part of a frikkin great simulation using 3-D "lasers"? (I've been feeling a bit fractal recently.)
Hey! This would make a great plot for a movie!!
-
I suppose when one looks at teraflops, the issue is that the video cards are massively parallel, so the applications must be written to take that into account.
Years ago I was doing experiments with simulating neural nets with sequential architecture. It would be a wonderful experiment to try the neural nets on a video card.
Rather than "Matrix" movie style screens, I wonder what it would be like to run a neural net app using VRAM, and then display a representative of the memory snapshot in high-res.
Anyway, so for my question. Probably 512mb or 1GB is sufficient for normal DVI-D display with heavy text usage. One should also look at the teraflops, but only if the application warrants it.
I'll get my new video card installed tomorrow, so hopefully I'll have some normal activity data to look at while using it.
-
Well...
I finally got the 2GB nVidia card installed with two monitors.
2560x1600 & 1920x1200, with a virtual screen, 4480x1600.
I got two 1080p, and one 720p videos running, and got up to about 603mb used. But, it took a lot of work to get up above 512MB.
So, I think I'll conclude that for ordinary, non-gaming use, 512MB, or perhaps 1GB would have been more than sufficient for memory capacity. I don't have any optimized 3D CAD or other apps here at the moment.
2GB wasn't that much more expensive than 1GB, but it doesn't appear to be necessary.
Multi-Cores?
So far I'm doing ok with only 48 cores.
So... I don't think I'll need to get 1024 CUDA Cores for normal work.
The question is whether I wish to try out massive parallel processing as suggested above, in which case, I can imagine the applications of a 1024 core machine... Perhaps 2 video cards for 2048 cores? Can I squeeze in more cards?
I had simulated neural nets in college using sequential processing, swapping to a 72MB (yes, a HUGE 72MB) Hard Drive, and it would take days to run the simulation.
I can only imagine how the net would have acted on a multi-core processor with lots of memory (VRAM and main memory if necessary).
The Matrix movie showed streams of funky characters on the screen.
What about doing multiprocessing on a graphics card, and allowing some of the memory to leak out into the bitmap area of high-res display. I could imagine most unique patterns showing up for monitoring the processing.
-
So, does this mean we are really all just part of a frikkin great simulation using 3-D "lasers"? (I've been feeling a bit fractal recently.)
https://en.wikipedia.org/wiki/Cellular_automata#CA_as_models_of_the_fundamental_physical_reality
Threshold automata have been invented to simulate neurons, and complex behaviors such as recognition and learning can be simulated.
https://en.wikipedia.org/wiki/Cellular_automata#Biology