A funny thing has happened every time we've set out to test polygon throughput using the Beyond3D test suite on Nvidia's Maxwell—and, more recently, Pascal—GPUs. Those chips have always delivered better performance in that test than their peak theoretical numbers would suggest. At least as far back as the GeForce GTX Titan X, we've suspected that Nvidia was using a technique called tile-based rasterization in its chips to increase their efficiency. Now, friend of TR David Kanter has applied a directed test of his own to Nvidia's Maxwell and Pascal chips to demonstrate that they are, in fact, using a form of tile-based rasterization. We won't spoil David's full article at Real World Tech for you here, but we will repost his video for your enjoyment.
Thanks to this investigation, we finally have a good explanation for why Maxwell and Pascal chips behave the way they do in our Beyond3D test results. We also have a compelling explanation, at least in part, for Nvidia's efficiency advantage over AMD GPUs, since—as David notes in his video—those chips still rely exclusively on immediate-mode rendering to do their thing. It'll be interesting to see whether Nvidia is more open about this approach in its whitepapers and architectural discussions going forward.