I didn't see a difference on the GV100. I have yet to test on the other cards. Here are the results:
Running ./jit
Run on (24 X 3326 MHz CPU s)
CPU Caches:
L1 Data 32K (x12)
L1 Instruction 32K (x12)
L2 Unified 256K (x12)
L3 Unified 12288K (x2)
***WARNING*** CPU scaling is enabled, the benchmark real time measurements may be noisy and will incur extra overhead.
Context:
ArrayFire v3.7.0 (37b7da45) (Backend: CUDA Platform: CUDA)
Device: Quadro_GV100 (7.0 v10)
|---------------------+---------------+---------+-----------|
| Test (jit/f32) 1 op | reletive diff | int | long long |
|---------------------+---------------+---------+-----------|
| dim0:1 | -0.0220 | 27231 | 26632 |
| dim0:8 | -0.0186 | 26832 | 26332 |
| dim0:64 | -0.0562 | 27009 | 25491 |
| dim0:512 | -0.0626 | 26851 | 25171 |
| dim0:4096 | -0.0601 | 27135 | 25503 |
| dim0:32768 | -0.0205 | 27152 | 26596 |
| dim0:262144 | -0.0090 | 28556 | 28298 |
| dim0:2097152 | -0.0081 | 56033 | 55580 |
| dim0:16777216 | -0.0199 | 262290 | 257069 |
| dim0:134217728 | -0.0310 | 1928646 | 1868893 |
| dim0:536870912 | -0.0384 | 7779973 | 7481318 |
|---------------------+---------------+---------+-----------|
|---------------------+---------------+---------+-----------|
| Test (jit/f32) 8 op | reletive diff | int | long long |
|---------------------+---------------+---------+-----------|
| dim0:1 | -0.0189 | 46598 | 45716 |
| dim0:8 | -0.0150 | 46604 | 45905 |
| dim0:64 | -0.0277 | 47098 | 45791 |
| dim0:512 | -0.0203 | 46911 | 45957 |
| dim0:4096 | -0.0223 | 47201 | 46150 |
| dim0:32768 | -0.0241 | 47558 | 46413 |
| dim0:262144 | -0.0130 | 49176 | 48538 |
| dim0:2097152 | -0.0086 | 77143 | 76482 |
| dim0:16777216 | -0.0160 | 282357 | 277842 |
| dim0:134217728 | -0.0274 | 1943859 | 1890669 |
| dim0:536870912 | -0.0486 | 7789002 | 7410695 |
|---------------------+---------------+---------+-----------|
|----------------------+---------------+---------+-----------|
| Test (jit/f32) 64 op | reletive diff | int | long long |
|----------------------+---------------+---------+-----------|
| dim0:1 | +0.0077 | 530187 | 534281 |
| dim0:8 | +0.0063 | 530706 | 534047 |
| dim0:64 | +0.0125 | 531338 | 537987 |
| dim0:512 | -0.0115 | 260390 | 257395 |
| dim0:4096 | -0.0177 | 261764 | 257131 |
| dim0:32768 | -0.0430 | 268395 | 256864 |
| dim0:262144 | -0.0166 | 261226 | 256893 |
| dim0:2097152 | -0.0116 | 288888 | 285542 |
| dim0:16777216 | -0.0156 | 814691 | 801955 |
| dim0:134217728 | -0.0284 | 5803226 | 5638558 |
| dim0:536870912 | +0.0076 | 8349142 | 8412499 |
|---------------------+---------------+---------+-----------|
|-----------------------+---------------+----------+-----------|
| Test (jit/f32) 256 op | reletive diff | int | long long |
|-----------------------+---------------+----------+-----------|
| dim0:1 | -0.0060 | 2679645 | 2663476 |
| dim0:8 | -0.0064 | 2680828 | 2663715 |
| dim0:64 | -0.0065 | 2684662 | 2667234 |
| dim0:512 | -0.0259 | 995418 | 969662 |
| dim0:4096 | -0.0338 | 1001301 | 967436 |
| dim0:32768 | -0.0348 | 999364 | 964597 |
| dim0:262144 | -0.0321 | 998454 | 966422 |
| dim0:2097152 | -0.0205 | 1018834 | 997992 |
| dim0:16777216 | -0.0182 | 2485035 | 2439760 |
| dim0:134217728 | -0.0322 | 19164272 | 18546961 |
| dim0:536870912 | +0.1427 | 25328483 | 28943381 |
|---------------------+---------------+---------+-----------|
All times are in nanoseconds