Graphics cards for GPU rendering?

[ellarien]

I want to replace my aged desktop PC (bought as a mid-range general-purpose home machine in 2007) some time this year -- though not immediately, because first I have to finish digitizing the family slide collection with a scanner that isn't Windows 8 compatible. I'd really like to play with GPU fractal rendering programs, but I know nothing about graphics cards except that the one I have (nVidia GeForce 8300 GS -- don't laugh too hard! ) is hopelessly outdated and underpowered. I don't want to spend a fortune -- no more than £1000 for the whole system including software and warranty but probably without a monitor -- but I want something more powerful and better cooled than my Core i5 laptop, so I'm probably looking at an entry-level 'gaming' machine even though I'm not a gamer.

So, how many CUDA cores and how much dedicated RAM do I need for a decent GPU rendering experience? Would something like the NVIDIA GeForce GTX 745 4GB DDR3 (384 CUDA cores) do the job? Do I even need a dedicated graphics card? My brother-in-law was trying to tell me that integrated graphics would work -- is that true?

One more question -- will a modern graphics card support a seven-year-old monitor (with DVI-D and D-sub connections), or am I going to need a new one?

Thanks for any insights you can provide -- even helpful search terms would be welcome!

! 



 

[claude]

first I have to finish digitizing the family slide collection with a scanner that isn't Windows 8 compatible.

Chances are the scanner might be Linux compatible - I inherited my dad's old printer when he upgraded Windows and no driver was available, it works fine with my Linux system.  So you could maybe try a Linux live boot system (CD, DVD, USB) that won't modify anything on your hard drive, and see if the scanner works.   Check the package list of the live system before downloading, needs GIMP and GIMP plugin for SANE (which is the Linux scanner system - you don't strictly need GIMP, but command line scanning is more awkward).

So, how many CUDA cores and how much dedicated RAM do I need for a decent GPU rendering experience? Would something like the NVIDIA GeForce GTX 745 4GB DDR3 (384 CUDA cores) do the job? Do I even need a dedicated graphics card? My brother-in-law was trying to tell me that integrated graphics would work -- is that true?

Integrated graphics would work fine for displaying the results of CPU rendering, and are increasingly more powerful, but for GPU rendering you really would want something much more powerful I think.  Also, you probably want to consider double precision floating point performance on the GPU, something not typically covered in game-oriented benchmarks (single precision floating point is pretty much the standard there).  You might need to look for science/engineering-oriented benchmarks.  Double precision offers much greater zoom depth than single precision floating point, which could be important depending on what you want to render.

Fairly standard point: it's probably not worth going for top-of-the-range, look for best price/performance ratio (unless you absolutely need maximum performance and are willing to pay over the odds for it).

One more question -- will a modern graphics card support a seven-year-old monitor (with DVI-D and D-sub connections), or am I going to need a new one?

It probably is possible to get a graphics card with DVI outputs still, though HDMI is increasingly the preferred option.  D-sub (analogue VGA if I understand correctly?) is obsolete by now I think.  Otherwise the motherboard will probably have some integrated graphics too, so you could render on your fancy GPU and display the results on the integrated output - might be a nightmare to set it up though.  Check the output connector specs before you buy.  Another thing to look into is multiple outputs on your graphics card, in case you eventually decide that 1 screen is not enough.


EDIT: found some fp64 double precision benchmarks (seems AMD is doing better these days, so unless you absolutely need CUDA then maybe broaden your search from NVIDIA alone)
http://www.tomshardware.com/reviews/geforce-gtx-780-performance-review,3516-28.html
http://www.theinquirer.net/inquirer/review/2162193/nvidias-gtx680-thrashed-amds-mid-range-radeon-hd-7870-gpu-compute

[3dickulus]

GeForce GTX 760 has 6 chips 192 cores each = 1,152 total with 2G ram was around $250 CDN, it came with DVI and a D-Sub adapter in the package, I have yet to exploit it's full capabilities, works reeeeaaaally well with Fragmentarium

Integrated gfx card will add $ and heat to mother board, a separate card can be cooled better and can be swapped out if you find it's not what you need. usually integrated gfx cards will be a "slimmed down" version that uses part of your system ram while a dedicated card will have it's own ram which is more likely to be tailored to the gfx chips needs, bus width , access speed etc.

I personally prefer building from the ground up, piece by piece, for a powerful gfx card you will also need a good power supply too.

[ellarien]

Chances are the scanner might be Linux compatible - I inherited my dad's old printer when he upgraded Windows and no driver was available, it works fine with my Linux system.  So you could maybe try a Linux live boot system (CD, DVD, USB) that won't modify anything on your hard drive, and see if the scanner works.   Check the package list of the live system before downloading, needs GIMP and GIMP plugin for SANE (which is the Linux scanner system - you don't strictly need GIMP, but command line scanning is more awkward).

Sadly not in this case. I did get my old Canon flatbed working with a Linux VM, but the slide scanner is a bit more specialized and a quick Google suggests there's no Linux support for that model. Anyway, I only have about a hundred slides to go ...

Integrated graphics would work fine for displaying the results of CPU rendering, and are increasingly more powerful, but for GPU rendering you really would want something much more powerful I think.  Also, you probably want to consider double precision floating point performance on the GPU, something not typically covered in game-oriented benchmarks (single precision floating point is pretty much the standard there).  You might need to look for science/engineering-oriented benchmarks.  Double precision offers much greater zoom depth than single precision floating point, which could be important depending on what you want to render.

Good point on the double precision, and one I hadn't thought of. Thanks!

[ellarien]

I personally prefer building from the ground up, piece by piece, for a powerful gfx card you will also need a good power supply too.

Thanks for your reply. I'm a little nervous about custom-building a system if there's an off-the-peg solution that will do what I want, though I suppose it ought to be possible with enough research.

[3dickulus]

it's actually rather simple only 8 or 9 components

case
power supply
mother board
hard drive(s)
cd/dvd reader/writer
ram
gfx card
and maybe an extra cooling fan or two if you are overclocking

motherboards usually come with (in Canada) instructions and the required mounting hardware and cables
and it's hard to plug them in wrong because they are all "keyed" to their respective plugs, generally, things only fit where they belong
you have a mechanically inclined nerdy friend who can help ?
it would be a good idea to have someone on hand with at least some experience for your first build
but don't let it intimidate you, if you can do a jigsaw puzzle you can assemble a computer

there's lots of vids on you tube, watch a few and you'll see it's not all that complicated.

 

[ellarien]

Just to round off the thread, what I eventually did was to buy a custom machine from a company that tailors to gamers, with a GTX 750Ti (2Gb, 640 cores). It's not the fastest thing on the planet, but it fit my budget and it does let me play with Fragmentarium. It turns out that even the integrated graphics on my laptop will do Fragmentarium up to a point, but the 3d rotation function doesn't work.

[3dickulus]

just on time for the competition