GT 730 specs comparison: CUDA Cores; 64/128bit; DDR3/GDDR5, MHz, Bandwidth, 1GB/2GB

I wonder why nVidia offers these type of configurations of it's graphics cards, in this case > GT730.

I know that bandwidth = Memory clock x Memory interface widthbut what about memory config or CUDA Cores? Where they come to play?

1. I wonder why 64 bit version has 2GB and not 1GB? (while 128 bit version has 1GB).
2. I wonder why GDDR5 version has 1GB and not 2GB? (while DDR3 version has 2GB).
3. I wonder why GDDR5 version has 64 bit and not 128 bit? (while DDR3 version has 128).

Also I wonder if the 64-bit 4GB performs same as the 128-bit 2GB card? Or the 64-bit 2GB version performs the same as the 128-bit 1GB card?

2 Answers

I wonder why 64 bit version has 2GB and not 1GB? >While 128 bit version has 1GB. 

This is almost certainly because the GPU has two memory controllers.

You get either two 64-bit channels capable of using 1GB of memory each, or one 128-bit channel capable of using a single 1GB memory area.

I wonder why GDDR5 version has 1GB and not 2GB? >While DDR3 version has 2GB. 

Probably because the extra bandwidth supplied by GDDR5 requires some internal multiplexing in the GPU which limits it to one memory channel. The real reason would only be known to GPU designers and Nvidia themselves.

Also I wonder if 64bit 4GB performs same as 128bit 2GB card? or 64bit 2GB perfors same as 128bit 1GB Card?

No. The "bitness" of the memory affects the effective memory bandwidth available. Look at the figure just below the memory interface width.

Graphics tasks tend to be incredibly memory bound, more memory bandwidth generally means it will perform better. Even then there may be tasks where a lower bandwidth card with more memory might perform better than a card with less memory but higher bandwidth.

You are down comparing the weeds amongst several (quite frankly) very poor cards and I seriously doubt you are going to see a massive performance difference between them.

As to why: product differentiation and because they can.

CUDA cores are calculated based on the number of multiprocessors present in the GPU as well as the number of ALU lanes and the width of the ALU lanes.

For my GeForce GT 730 (DDR3, 64-bit version):

(2 multiprocessors * 192 64-bit ALU lanes) = 384 CUDA cores

The memory bandwidth is also calculated based on the number of multiprocessors on the GPU:

(2 multiprocessors * 800 MHz) = 1600 MHz effective

On this device, single-precision floating point operations are 24x faster than double precision. This may vary depending on the bus width and number of multiprocessors, but I don't currently have anything else to compare to.