On AMD cards, FahCore 17 is about 10 times faster than the old GPU cores, and on Nvidia it's about twice as fast. This is mainly due to its OpenMM 5.1 base, which contains many optimizations which deliver a significant speedup. One optimization in particular that we are waiting for is CUDA JIT, a just-in-time compiler that Nvidia may be introducing into its drivers in the coming future. Not only will this technology allow us to offer support for the CUDA platform with FahCore 17, but the JIT compiler is likely to deliver a massive speedup. For the time being, we continue to work at finding additional optimizations on our end. We have also successfully tested FahCore 17 with extremely large proteins (500,000+ atoms), which are on par with the ones used by "bigadv" CPU projects.
To run FahCore 17, you need a Fermi GPU or better and Windows or Linux, or a AMD HD5000 or better and Windows. It also currently requires proprietary drivers from these vendors. You can test FahCore 17 by adding the "client-type = advanced" setting into the extra core options in the V7 client, as in the Configuration FAQ. Another excellent resource is the GPU FAQ which describes why GPUs are so helpful to us.
We'd like to thank all the alpha testers on FreeNode's #fah IRC channel, as well as the beta testers on foldingforum.org, who have all helped us bring the core to this point!