Unified Virtual Memory Stripped from Maxwell

[Chalenge]

Sadly, the big benefits of the promised new methods will not be available until 2016. Worse, though Nvidia acknowledges that PCIe just is not fast enough, it will only be putting NVlink in servers.

Apparently, the market will not support a big technology jump at this time (it's too expensive).

http://www.anandtech.com/show/7900/nvidia-updates-gpu-roadmap-unveils-pascal-architecture-for-2016

[MrRiplEy[H]]

It seems like NVlink is required only in multi-gpu setups so it's no wonder they left it out. The user base for high-end multi gpu setups is marginal at best.

[Chalenge]

NVLink was never part of Maxwell. The unified virtual memory was to be the biggest game changer yet, but delaying it will possibly help to bring the end-market price down. The problem is that until that happens we are stuck with limited memory frames.

NVLink will help to change that while also keeping the costs down, but even if it could be introduced today it would be too expensive.