Manycore

Previous work

Interface Pyclaw with CUDACLAW to make CUDACLAW a legitimate child of the Clawpack family. This will involve cleaning up and refactoring the existing code.
Automatic generation of CUDA kernels for Riemann solvers from an appropriate high level description. One of the questions here is the nature of the high-level language. A Matlab-like language perhaps?
Automatic generation of kernels for setting up boundary and initial condition, source terms, etc. needed to set up problem instances. We really should have it as a goal to make it relatively straightforward to run CUDACLAW: no need for a user to know CUDA to set up a simulation. The system won't likely get much traction otherwise.
Multi-GPU issues. 3D and high-res 2D simulations will inevitably require the data to be spread on multiple GPUs.
Using CUDACLAW as a library to, for example, advance a subgrid for a few time steps and return to the CPU. Such a facility might be useful in a variety of contexts including a CPU-based patch-oriented AMR where the GPU is used as an accelerator for certain subtasks. The question here is how should the API look like for such a capability? Coordination with the AMR working group will be needed here.
Visualization/steering. User specification of (realtime) visualization. Interactive steering in the context of GPUs. We obviously need to coordinate with the Viz/Steering working group.
OpenCL and other languages for targeting MIC and other SIMD targets, as per the earlier thread of Andy and Nathan.