tpsieve-cuda slows down with increasing p
 

Since I switched to using tpsieve-cuda at p > 21T for the 495000-50000 sieve, I have noticed that its throughput appears to go down with increasing p: 355M p/sec at 27T down to 318M p/sec at 430T based on wall clock rather than CPU time. Throughput based on CPU time shows a similar decrease.

tpsieve throughput increased with increasing p which is what I would expect based on other sieving programs. Is the behaviour of tpsieve-cuda to be expected? Note that I have dedicated a CPU core to the sieve, although usage of the core has increased from 0.86 at 27T to 0.91 at 430T.