Which cards and compute capabilities are required to fully utilize CUDA 5's features [closed]

Question

We just received the stable version of CUDA 5. There are some new terms like Kepler and ability of using MPI with better performance, and running the same card with 32 applications at the same time. I am a bit confused though and looking for the answers of such questions:

• Which cards and compute capabilities are required to fully utilize CUDA 5's features?
• Are new features only available for Kepler architecture, like GPUDirect, Dynamic Parallelism, Hyper Q and Dynamic Parallelism.
• If we have Fermi architectures, what are the benefits of using CUDA 5. Does it bring benefits other than ability of using NSight at Linux and Eclipse. I think the most important feature is ability of building libraries?
• Did you see any performance improvements by just passing from CUDA 4 to CUDA 5. (I got some speed ups at Linux machines)

I found out some documents like

• http://developer.download.nvidia.com/compute/DevZone/docs/html/C/doc/Kepler_Compatibility_Guide.pdf
• http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf
• http://blog.cuvilib.com/2012/03/28/nvidia-cuda-kepler-vs-fermi-architecture/

However a better, short description may make our minds clearer.

PS: Please do not limit the answer to the questions above. I might be missing some similar questions.

Answer 1

Compute capability 3.5 (GK110, for example) is required for dynamic parallelism because earlier GPUs do not have the hardware required for threads to launch kernels or directly inject other API calls into the hardware command queue.

Compute capability 3.5 is required for Hyper-Q.

SHFL intrinsics require CC 3.0 (GK104)

Device code linking, NSight EE, nvprof, performance improvements and bug fixes in CUDA 5 benefit Fermi and earlier GPUs.