Anyone know when GTX 970/980 with 8GB vram may come out?
Scenario question: If the card can only effectively use 4GB, but they give it 8GB, and the game tries to use 8GB, will the card use it effectively? Or will it flounder?
Also, do you have any links or directions to look for a good explanation about "effective" use of VRAM in gfx cards? I understand addressable limits and such, but not why a card couldn't use more RAM. Unless it is simply that the card only has enough processing power to use X amount.
It's a matter of memory bandwidth, the GPU buffers frames in the memory continuously and pulls them out to send to the display continuously, so there is a huge amount of information transfer going in and out across the memory bus at all times. The memory bus has a certain maximum bandwidth (bus width times effective frequency, 256-bit x 7GHz = 1.792Tbits/s = 224GB/s) and so it can only facilitate so much communication. If the memory bandwidth is being maxed, then the GPU simply won't be able to store or retrieve frames any faster than the memory bus is capable of handling per second, which translates into a framerate limitation. Any extra frames that the GPU is capable of generating are simply lost, as it has nowhere to buffer the frame since the memory bus cannot transfer any more data to the memory than it already is. If you are running at very high resolution with very large frames to buffer, and you are limited by memory bandwidth, you'll have a lower framerate than running at lower resolution (again, assuming everything is limited by memory bandwidth). The memory bus has a maximum data transfer rate, and larger frames contain more data, so not as many of them will fit across the bus per second. At this point adding extra capacity won't matter, because the GPU can't fill any more of the VRAM capacity with frames, data transfer to the memory cannot go any faster. It's like a congested road, it doesn't matter how big the parking lot at the other end is, if the road only has one lane, there is a maximum limit to how many cars can come and go in an hour (assuming everyone respects the speed limit). So... no point in expanding that parking lot beyond a certain size, until you add some lanes (wider bus) or crank up the speed limit (higher frequency).
This is very application dependent though; if you are using the GPU as a GPU, the memory is used for frame buffering, and the above applies. Things are a bit different if you are using the GPU as a co-processor for accelerating CPU computations (in a word, GPGPU). In this type of usage it is not using the VRAM for storing frames and retrieving them constantly and as fast as possible. It does store a large amount of information that it needs for doing its work, but in this case much more memory can be utilized effectively, because it is not trying to access massive amounts of it at once, the data transfer and communication is comparatively very small compared to frame buffering. It just stores a lot of information, and accesses a little bit at a time. Just what it needs at any given point for its computations. That's why for example you see a Quadro K6000 with 12GB of VRAM on a 384-bit 6GHz bus (288GB/s) and it uses all of it no problem.
Of course, you don't even need to go into memory bandwidth for this situation, it's a simple matter of GPU raw computation power. A GTX 980 is not going to be generating enough frames in gaming to fill up 8GB of VRAM anyway, and the GPGPU capabilities are severely crippled in GeForce cards (artificially) so it's a very bad option for that. Maybe it makes sense when you think about two-way or three-way SLI (since the memory capacity is not combined) but then again, you run into bandwidth issues in that case...
Create an account or sign in to comment
You need to be a member in order to leave a comment
Create an account
Sign up for a new account in our community. It's easy!
Register a new accountSign in
Already have an account? Sign in here.
Sign In Now