You're right. GPU on any ARM SoC will never run full-fledge CUDA code (except Maxwell I guess), but the benefit of dynamic GPU resource allocation is still usef

You're right. GPU on any ARM SoC will never run full-fledge CUDA code (except Maxwell I guess), but the benefit of dynamic GPU resource allocation is still useful if we believe computing resource should be as flexible as storage in cloud environment