Sign in with
Sign up | Sign in
Your question

Need Advice: Budget GPU computer for scientific calculations

Last response: in Graphics & Displays
Share
October 9, 2011 8:47:15 PM

Hello,

My question is what is the cheapest GPU machine I can buy capable of double precision floating-point calculations. I would like to use this machine to test my software's compatability before stepping up to a much more expensive machine running multiple Nvidia Tesla 2050 GPUs.

The focus of my project will be correlation analysis on large climate databases. I need to use a GPU to get the processing time down. My project uses the R programming language and will run on a LINUX OS.

I am not very familiar with computers and I am not sure which options are neccessary and provide a good return on investment. Do I need an Intel Core i7 or will a Core i3 suffice. How much cooling do I need? What size power supply? How much memory and what type? Hard drive or solid state storage?

Any help is greatly appreciated!

APPROXIMATE PURCHASE DATE: (this week or next)) BUDGET RANGE: (e.g.: USD $500-$700) Before / After Rebates

USAGE FROM MOST TO LEAST IMPORTANT: (scientific/statistical computing using R programing language (R-GPU, RPUD)

CURRENT GPU AND POWER SUPPLY: (None yet, currently researching purchase options.)

OTHER RELEVANT SYSTEM SPECS: (need GPU to be capable of double precision floating-point calculations. Others have recommended Nvidia GTX 460 or better)

PREFERRED WEBSITE(S) FOR PARTS: (http://www.avadirect.com/desktop-pc-configurator.asp?PR...)

PARTS PREFERENCES: (plan to use Nvidia GTX 460 or better)

OVERCLOCKING: No / Maybe SLI OR CROSSFIRE: No

MONITOR RESOLUTION: (low resolution, display quality not an important factor)

ADDITIONAL COMMENTS: (system will be running for long periods of time. Days or weeks crunching large data sets)

-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
October 10, 2011 2:27:24 AM

All dedicated modern GPUs will calculate floats. For $700, you could get a pair of GTX 570s, although I don't know anything about using them with the R language; only openCL for gpu computing.

A GTX 460 is only about $140 for the 1GB version.
m
0
l
a b Î Nvidia
October 10, 2011 3:10:47 AM

I believe R can be used in CUDA, though I have only done CUDA C. Since this is all about the GPU, an i3 would do fine. All the CPU is doing are the serial tasks and driving the GPU.

Get the 470/570 or 480/580, due to the architecture of the 460/560, they are difficult to fully utilize in CUDA.

Remember, the gaming cards (4xx and 5xx, the non-tesla and non-quadro) cards have crippled double precision - 1/4 that of their compute counterparts. However you might be memory bandwidth bound anyway, depends on the code..
m
0
l
Related resources
October 14, 2011 12:54:09 PM

jryan388 said:
All dedicated modern GPUs will calculate floats. For $700, you could get a pair of GTX 570s, although I don't know anything about using them with the R language; only openCL for gpu computing.

A GTX 460 is only about $140 for the 1GB version.


Thanks for the ideas!
m
0
l
October 14, 2011 1:02:09 PM

EXT64 said:
I believe R can be used in CUDA, though I have only done CUDA C. Since this is all about the GPU, an i3 would do fine. All the CPU is doing are the serial tasks and driving the GPU.

Get the 470/570 or 480/580, due to the architecture of the 460/560, they are difficult to fully utilize in CUDA.

Remember, the gaming cards (4xx and 5xx, the non-tesla and non-quadro) cards have crippled double precision - 1/4 that of their compute counterparts. However you might be memory bandwidth bound anyway, depends on the code..


Thanks for the advice! I really don't know much about the differences in the cards. My goal is to test my code to see if it functions well in a GPU envionment on the most economical card possible, then step up to a Tesla 2050 or 2070 for the heavy lifting.

If there is a card in the $200-$300 price range that is a better choice than the GTX 460 I would be interested. You mention 470/570 or 480/580 as a good choice but that there are limitations as they are gaming cards. Is there an affordable Quadro or TESLA card near my price range?

One last question you mention memory bandwidth restrictions. What should I be looking out for? Is it memory on the GPU card or CPU memory?

Thanks,

Allon
m
0
l
October 14, 2011 10:38:06 PM

"Affordable" and "quadro/tesla" do not belong in the same sentence.
From a quick check on tigerdirect,

Tesla C2070, 6GB ram and 448 cores, is about $2.5k.
Quadro 6000, 6GB ram and 448 cores, is about $4k.
GTX 580, 1.5GB ram and 512 cores, is around $450.

The reason they are so expensive is that they are optimized for intensive use. I believe they have ecc memory and some other features. They have drivers written specifically for cad, modeling, rendering, etc. programs that let them run much faster than GeForce cards.
m
0
l
October 15, 2011 1:28:55 PM

jryan388 said:
"Affordable" and "quadro/tesla" do not belong in the same sentence.
From a quick check on tigerdirect,

Tesla C2070, 6GB ram and 448 cores, is about $2.5k.
Quadro 6000, 6GB ram and 448 cores, is about $4k.
GTX 580, 1.5GB ram and 512 cores, is around $450.

The reason they are so expensive is that they are optimized for intensive use. I believe they have ecc memory and some other features. They have drivers written specifically for cad, modeling, rendering, etc. programs that let them run much faster than GeForce cards.


Thank you for clarifying.

-Allon
m
0
l
!