• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

NVIDIA Teasing something for 29th April

ToTheMax said:
Wouldn't it be neat if they released a dual gpu 1090 with a shared memory pool architecture, meaning no need for SLI shenanigans. One can dream...
Click to expand...

Even with a shared memory pool you still need SLI and will get the same scaling performance, or actually somewhat less. If 2 8GBU could share their memory then you would have the advantage of 16GB memory rather 8GB duplicated, but the bandwidth would have to be shared between 2 GPUS so the effective bandwidth per GPU would be halved. In a memory BW limited situation performance would drop.
Beyond that, just because 2 GPUs access the same memory doesn't mean there is any differences in the requirement for the GPU to coordinate rendering and share the load. The easiest way to do that is still to do AFR. You can imagine all kinds of complex situations where the screen is split in to tiles and like a chessboard and the GPUs render each tile. But the problem is that some resources are required on multiple tiles o you loose all the advantages of cache coherency and a lot fo the very smart optimization that exist in GPUs to do things like running the same shader over thousands of pixels becomes somewhat dilutes. So actually you need a far more complex SLI system with far more integrated developer work where some sub-tasks are rendered individually by each GPU and then finally assembled. This is highly complex to optimize. Or we can just carry on using AFR.
 
D.P. said:
Even with a shared memory pool you still need SLI and will get the same scaling performance, or actually somewhat less. If 2 8GBU could share their memory then you would have the advantage of 16GB memory rather 8GB duplicated, but the bandwidth would have to be shared between 2 GPUS so the effective bandwidth per GPU would be halved. In a memory BW limited situation performance would drop.
Beyond that, just because 2 GPUs access the same memory doesn't mean there is any differences in the requirement for the GPU to coordinate rendering and share the load. The easiest way to do that is still to do AFR. You can imagine all kinds of complex situations where the screen is split in to tiles and like a chessboard and the GPUs render each tile. But the problem is that some resources are required on multiple tiles o you loose all the advantages of cache coherency and a lot fo the very smart optimization that exist in GPUs to do things like running the same shader over thousands of pixels becomes somewhat dilutes. So actually you need a far more complex SLI system with far more integrated developer work where some sub-tasks are rendered individually by each GPU and then finally assembled. This is highly complex to optimize. Or we can just carry on using AFR.
Click to expand...

I thought AFR had issues though in DX12 where prior frames are referenced to render the current frame?
 
AFR breaks down in a lot of situations - but its generally superior to other potential methods.

The only thing that is superior and potentially a realistic implementation really is with stuff like DX12 explicit multi adapter where the developer has more control over farming out the processing in a manner they know will be most efficient with least duplication or dependencies and unfortunately a lot of developers seem to drag their heels over stuff like that and/or lack the imagination to implement for best effect until one of the big names gives it a go.
 
You must log in or register to reply here.
Back
Top Bottom