That's a great point about --maxmem.
I updated my
optimizer. New hypothetical numbers for a t70 on the C248 remainder of M1217:
147K curves at B1=3e9 B2=6.4e12
Stage 1 (on GPU) takes 120 seconds/curve and Stage 2 (on CPU) takes 285 seconds/curve
Peak memory was reported as 3244MB
The GPU would complete a batch every 2.5 days which is a little more than I'd like.
204 GPU-days + 484 CPU-core-days
111K curves at B1=4.3e9 B2=12.9e12
GPU 165s / curve, CPU 370 s / curve
Peak memory was reported as 3249MB
217 GPU-days + 475 CPU-core-days
260K curves at B1=1.2e9 B2=22.4e12
GPU 48s / curve, CPU 474s /curve
Peak memory was reported as 6596MB (so I might run into trouble with running more than a few instances)
GPU batches would complete every day which is nice.
144 days GPU-days + 2100 CPU-core-days
The 3rd cost slightly more (10 CPU-core-days = 1 GPU day) but finishes faster. Sadly the long batches from the first option mean I probably can't actually do this (I may try one batch just to validate numbers).