Unrolling a memory-bound loop at 32 versus 96 registers per thread

An SM has 65,536 32-bit registers and supports up to 2048 resident threads (64 warps). This kernel is memory-latency-bound (each thread stalls frequently waiting on global memory). Version A uses 32 registers/thread. Version B aggressively unrolls the inner loop, reducing loop-overhead instructions per element but raising usage to 96 registers/thread. What is the most likely outcome for Version B relative to Version A?

9 of 46 in GPU / CUDA