Vlada011 wrote:
What will be with SLI when Pascal show up with 1TB Bandwidth? NVIDIA think to keep him or to remove as AMD?
And what will be with bigger 1TB Bandwidth?

Pascal-based NVLink is anticipated to achieve peak transfer rates up to 80GB/s, nowhere near 1TB/s.

SLI allows up to 1GB/s, plus it can claim unused lane-dedicated bandwidth over the PCIe bus if needed. (And in case anyone asks, CrossFire apparently achieves up to around 900MB/s.)

Understand that games and applications are typically written for single-processor platforms - desktops, laptops, etc. Each GPU processes and calculates and renders and grinds out its own workload independently - inter-GPU communication is mostly needed only to coordinate GPU activities at a higher level. So not a lot of bandwidth is really needed. AMD's "bridgeless" CrossFireX doesn't even bother to use a physical link, it's just more packet traffic dumped onto the PCIe bus.

But GPU-based supercomputing involves bucketloads of GPUs crunching in parallel towards a common task. The processing efficiency of each GPU component tends to be increased when more complete information is made available, meaning that in a perfect world every single one of those dozens or thousands of GPUs has instantaneous access to the data every other one of those GPUs is calculating. Higher efficiency requires higher bandwidths, a little SLI bridge could become a tight bottleneck because it forces "simultaneously parallel" processes to feed through a long priority queue. Notice that most enterprise-class workstation GPU cards hardly even have a video output, they are instead stuffed with ports which link to other workstation GPU cards.

Admittedly, prosumer cards like Titans do blur the line a little, their upper potential might be constrained by little SLI bridges. The only thing holding them away from the server racks is deliberate lack of DFPP. But it turns out that games (so far) are really lousy at utilizing multi-CPU and multi-GPU hardwares, there are still many instances of games performing better or achieving higher fps when HyperThreading is deactivated or CF/SLI bridges are broken. Even if games can almost barely keep up with desktop hardware today, they certainly won't perform any better if given server hardware tomorrow - DX12 promises to deliver tools which can better balance and optimize multi-GPU loading, but it'll take years before they're properly used and ye olde SLI bridge is looked at as an obsolete interface.