Increase performance of CUDA code past 0.9 TeraOPS #36
Description
see
Several orders of magnitude slower than optimum (4x higher GPU allocation), multi GPU, Assembly optimization....
- 1 hour / quadrillion runs
- 2^27 search/sec around the 45 bit space (2^32 search/30-sec)
- 0.3 TeraOPS/sec (the goal would be at least 30 TFLOPS level - the RTX-A6000 FP8 maxiumum)
On an RTX-A6000 running CUDA C code at a 55% GPU saturation (24% TDP) we are getting around 3370 seconds per bit from bit 44 to 46 which is around 64-9 = 55 quadrillion search space. If we divide 3370 by 3600 we 0.94 h/quadrillion. Now we are checking only odd numbers so performance is double this.
Throughput
(bit 46 max - bit 44 max) = 55831878266636 / 2(odd optimization) = 27915939133318 runs
27915939133318 / 60h = 465265652222 runs/hr = 129240459 runs/sec = 2 * 2^27 = 2^28 search space / sec
Operations
We do on average one shift left and one shift right + an add - which averages to 1.5 operations * 1500 average path = 2250 ops/run
129240459 runs/sec * 1500 path = 193860688426 rough iterations/sec
With an average path of 1500 per run we are doing around 2250 ops/run.
operations/sec = 129240459 runs/sec * 2250 ops/iteration = 290791032639 IPS or 290791 MIPS or ~ 290 GigaFLOPS FP0 or .3 TerraFLOPS