Thread: AVX-ECM
View Single Post
Old 2020-11-03, 02:50   #67
bsquared
 
bsquared's Avatar
 
"Ben"
Feb 2007

64418 Posts
Default

Quote:
Originally Posted by VBCurtis View Post
When GMP-ECM is faster for Stage 2 with a higher B2, there is only advantage from using it. Save time, get more work- what's not to like?
If it were comparing curve-for-curve, sure, but it's not that straightforward. avx-ecm stage 2 may be slower and smaller B2, but it is doing 8 curves at once in that time.

Expanding on the example from before, with B1=7M on 2^1277-1:
avx-ecm stage 1: 78 sec for 8 curves, 9.75 sec/curve
avx-ecm stage 2: 55 sec for 8 curves, 6.87 sec/curve

gmp-ecm stage 1: 60 sec for 1 curve
gmp-ecm stage 2: 32 sec for 1 curve

Now, each avx-ecm stage 2 curve is 2.8 times less likely to find a factor, but if you had 64 seconds to spend doing something, would you rather run 2 gmp-ecm stage 2 curves or 8 avx-ecm stage 2 curves? Even with each avx-ecm being 2.8 times less likely to find a factor, you'd be better off.

Quote:
Originally Posted by VBCurtis View Post
The difficult tradeoff would be on larger numbers, when the avx-ecm stage 2 is likely faster than GMP-ECM but a smaller B2. That leads to some need for study.
The same math as above applies. I think you are better off running pure avx-ecm until time/curve * probability-multiplier is greater than gmp-ecm's stage 2 runtime.

It gets interesting with generic inputs, when the throughput advantage of avx-ecm is not as great. Then, I think it makes sense to run avx-ecm stage 1 followed by 8 gmp-ecm stage 2's. Then you get all of the probability advantage of gmp-ecm stage 2 but with avx-ecm's higher stage 1 throughput. I am putting together some data on this case.
bsquared is offline   Reply With Quote