Is there a "fpga" style configurable gpu shader core or cpu pipeline in the future?
For example, when software A is loaded, cores become 1 adder pipeline + 2 multipliers pipeline, when software B is loaded then cores become 2 adders and 1 multiplier.
Unused adders or multipliers are evaluated as some local cache or completely shut down.
So a non-optimized game is not an issue, skewed benchmarks are not a problem anymore, maybe some combinations use less power too?
Maybe an intact fpga chip embedded in gpu so we get a second opencl compute device per gpu, that adapts automagically to any incoming code block?