L1-resident compute: 9.93 ns/block (14.5 GB/s per core if memory were free)  [-3.66508e+08]
v1       pf= 0 chunk= 16 :  101.3 GB/s
v1       pf= 0 chunk= 32 :  103.1 GB/s
v1       pf= 0 chunk= 64 :  104.8 GB/s
v1       pf= 0 chunk=128 :  100.3 GB/s
v1       pf= 4 chunk= 16 :  120.9 GB/s
v1       pf= 4 chunk= 32 :  124.6 GB/s
v1       pf= 4 chunk= 64 :  124.2 GB/s
v1       pf= 4 chunk=128 :  117.8 GB/s
v1       pf= 8 chunk= 16 :  131.0 GB/s
v1       pf= 8 chunk= 32 :  138.9 GB/s
v1       pf= 8 chunk= 64 :  141.4 GB/s
v1       pf= 8 chunk=128 :  135.0 GB/s
v1       pf=16 chunk= 16 :  147.0 GB/s
v1       pf=16 chunk= 32 :  159.7 GB/s
v1       pf=16 chunk= 64 :  163.8 GB/s
v1       pf=16 chunk=128 :  157.7 GB/s
v1       pf=32 chunk= 16 :  152.4 GB/s
v1       pf=32 chunk= 32 :  166.8 GB/s
v1       pf=32 chunk= 64 :  174.8 GB/s
v1       pf=32 chunk=128 :  169.1 GB/s
rawread  pf= 0 chunk= 16 :  183.3 GB/s
rawread  pf= 0 chunk= 32 :  181.1 GB/s
rawread  pf= 0 chunk= 64 :  196.7 GB/s
rawread  pf= 0 chunk=128 :  194.8 GB/s
KB_DONE