Hello!
Our company develops a software, which depends currently on CPU power using Linux & jpeglib to read/scale/write JPEG images. Depending on situation, there can be dozens of JPEG images per second scaled to different sizes.
Is there any estimate if compared to jpeglib and for example a 2*Dualcore Xeon system, how much efficiency for example a single NVIDIA Tesla C870 card would provide?
A 2*Dualcore Xeon 2,1GHz system can handle up to 80 JPEG read/scale/write events per second.
I suppose, that JPEG it is better to realize on CPU as, it is fast (small amount of operations per pixel) and a host will stand whan device will be work.
Yes, I agree that reading JPEG-file and making the assumption that it’s really a JPEG-file is better on CPU.
But my guess is that using TESLA for hunders of scalings per second is more effective than using jpeglib and raw CPU power.
Nobody has yet written a JPEG implementation using CUDA as far as I’m aware, but it would certainly be possible.
The DCT/IDCT, subsampling and color conversion steps are easily parallelizable and could generate large speedups with CUDA. The entropy compression/decompression stages would be more difficult.