Video memory buswidth vs video memory Bandwidth

Question

My current video card (9600GT) is dying and I'm searching for a new video card. Between acquiring my current one and now, I got a lot more knowledge about hardware and I want to use that to pick my new card. So I decided to not just buy some popular card blindly, but to search for a card able to handle my hardware requirements.

I searched the specs at the NVidia site for the GT640 and was confused by the memory section and some questions raised. My current card's memory bus width is 256bit and has 1GB of memory. I checked Google about the importance of bus width. And all the links basically said the same 'The higher the number the more potential simultaneously traffic can be transferred'. This was already clear to me, yet there are currently a lot of new cards which are considered better than my current one with a lower bus width.

To go in more detail about my question I copied the memory info from the NVidia site:

                          GT 640        GT640 GDDR5
Memory Specs:
Memory Clock              1.8 Gbps      5.0 Gbps
Standard Memory Config    2048 MB       1024 MB
Memory Interface          DDR3          GDDR5
Memory Interface Width    128-bit       64-bit
Memory Bandwidth (GB/sec) 28.5          40.0

What puzzled me is that the Memory Bandwidth seems to me the most important part, yet the lower bus width has the higher 'performance'.
Is this due to the fact the memory interface is GDDR5 and is therefore able to have a higher memory clock speed (5Gbps)? If I am to buy a new video card, should I check the bus width? Memory clock? Bandwith? Amount of memory?

My current card ahs 1GB memory, so I was searching for a 2GB memory card, but now I'm not so sure any more whether that is really 'better'.

My main question:
To me it seems that memory performance is made up by the combination of bus width and frequency. Is this true? If yes, why are there so many sites telling me I need to get a card with a high bus width? If no, then what IS important when it goes about memory performance on a video card.

NOTE: The memory bandwidth is (almost) never displayed on vendor sites. How can I determine which card is better without knowing the bandwith?

Answer 1

Bus width is important, but only when you compare the same memory technology. GDDR is fundamentally different from DDR, it trades latency for bandwidth.

GPUs are data-parallel by nature, they benefit from memory architectures that can fetch large blocks of data at a time when they begin a task on a particular dataset, not so much small randomly distributed memory in as few cycles as possible. CPUs tend to want the latter to prevent stalls waiting for memory, so DDR is usually the domain of CPUs and GDDR for dedicated GPUs.

GDDR memory with a smaller bus-width is capable of outperforming DDR memory because the timings on GDDR allow for much greater clock speeds (at the expense of latency). It is not a fluke that the 64-bit bus width GDDR5 memory has greater throughput, but at the same time 40 GiB/sec bandwidth is laughable for GDDR5. If you saw that card up against a similarly clocked 384-bit GDDR5 card, it would be night and day.

Dedicated GPUs with DDR memory exist only in the very lowest-end value segment, DDR memory is cheap, but it actually hinders GPU performance much the way that using GDDR memory on a latency-sensitive CPU is less than ideal.

You definitely want the GDDR model, memory bandwidth is extremely important for real-time graphics such as games. Shadows and other effects eat memory bandwidth like there is no tomorrow. GPGPU applications do not benefit as much from GDDR, but I highly doubt you are in the market for such a low- to mid-range GPU to crunch numbers.

Update:

Is it the case that memory bandwidth usually is not published? I have not regularly looked at the consumer part of vendor sites in years.

Barring actual bandwidth numbers, the prime thing that bandwidth affects is fillrate. If bandwidth is not listed, texture/pixel fillrate is a pretty good indication. Pixel fillrate is what effectively limits your resolution and traditional (e.g. MSAA) anti-aliasing capabilities at playable framerates. Newer anti-aliasing techniques are based on image post-processing and move a lot of the burden from memory to compute.

Likewise, shadow quality is limited by both pixel fillrate during construction of shadow maps and texel fillrate during the application of shadows. If you are concerned about image quality, then fillrate is generally your number one consideration and that is calculated based on the number of ROPs and memory bandwidth. Neither one of those details may be listed by a vendor, but fillrate generally is listed in one form or another.

In the absolute worst case, Wikipedia has lists of very detailed hardware specs. for all major AMD, NV and Intel GPUs that exceeds the information you will find on the vendors' consumer sites.