https://hgpu.org/?p=5652
Quantifying NUMA and contention effects in multi-GPU systems