https://forums.developer.nvidia.com/t/blockchain-drivers/110847/12

We’ve confirmed that the performance drop is due to the size of the DAG exceeding the total on-chip TLB capacity on the Pascal GPU. As a result, there is an increased number of TLB misses, which affects performance. Because the TLB is a fixed capacity hardware resource, and the ETH algorithm accesses the DAG randomly, we don’t believe there are any software optimizations that could reduce the TLB miss rate.

Nvidia seemed to acknowledge the issue however they passed it off as a "hardware" problem. Really a bummer since that means 99% they won't release a driver fix; not to mention Nvidia probably hates crypto miners since angry shareholders opened a class action lawsuit regarding GPU sales