Is there an easy way to do the mapping of programs onto a heterogeneous system? It seems like most of the parallel interfaces we've used are tailored to specific contexts (e.g. CUDA for the GPU) and there is no way to leverage different resources without using multiple different interfaces (ouch).
solovoy
Is there any possibility that instead of changing software "in real time" based on available hardware, we change the hardware "in real time" to match the software requirement? Or in another word, are hardware architects trying to create computing units that can be installed in real time?
yuel1
@solovoy FPGAs could in theory do this. However, programming FPGAs are hard to program, computationally expensive to program (synthesize, sort of like compiling) and not as dense in terms of compute/unit area as composed to fixed function units.
Is there an easy way to do the mapping of programs onto a heterogeneous system? It seems like most of the parallel interfaces we've used are tailored to specific contexts (e.g. CUDA for the GPU) and there is no way to leverage different resources without using multiple different interfaces (ouch).
Is there any possibility that instead of changing software "in real time" based on available hardware, we change the hardware "in real time" to match the software requirement? Or in another word, are hardware architects trying to create computing units that can be installed in real time?
@solovoy FPGAs could in theory do this. However, programming FPGAs are hard to program, computationally expensive to program (synthesize, sort of like compiling) and not as dense in terms of compute/unit area as composed to fixed function units.