Interestingly, Yves provided some calculations on power efficiency for the sieve on Discord. Using a 7950X for comparison, the crossover for power efficiency came out a bit under 8000T. So this eems likely to come out close to that calculation.
____________

Range(T) Found Removed Efficiency 2000 107600 24339 6,7260 3000 61582 13921 3,8470 4000 43549 9801 2,7085 5000 33267 7225 1,9966

[...]
Looking how efficiency drops, it's possible that shutdown sequence will be initiated after 7000P or 8000P.

Many thanks for going to the trouble to quantify matters, Yves.

Hopefully there is a bit more time for the dinosaurs of the group to crunch for a little while longer. Foot on the gas pedal on a 24/7 basis.

I particularly like the target specified (ex 7000T, 8000T) and which anyone can easily monitor from the project task page.

- "Chip vs chip" is compared, i.e. runtime of one GPU is compared with runtime of one CPU running all threads in most efficient mode;

An optimal sieving depth is a point where checking candidates for primes is faster than removing them by sieving. Determining optimal sieving depth for CW sieve is very difficult because sieving is done on GPU, but crunching on CPU. We have to compare incomparable things with completely different architecture. An answer will depend on many factors - what and how we'll compare. For example,

- Should we compare runtime of single core or CPU chip as a whole?
- Which CPU and GPU should be used for benchmarks? Should we use common desktop models or, since new CW candidates will be tested in the future, use top AVX-512 models because they'll be common after few years?

This leads to lot of possible combinations. Some of them will give an answer "Yes, go on, sieving it still 20x effective", some of them will give "Stop immediately, it's already oversieved".

Finally, following rules were chosen:

- "Chip vs chip" is compared, i.e. runtime of one GPU is compared with runtime of one CPU running all threads in most efficient mode;
- "Almost topmost" CPU and GPU were chosen for benchmarks - Ryzen 7900X (AVX-512, running 2 tasks 6 cores each) and 4070 TI.

This gives following efficiency table:

Range(T) Found Removed Efficiency ------- taken from real data -------- 2000 107600 24339 6,7260 3000 61582 13921 3,8470 4000 43549 9801 2,7085 5000 33267 7225 1,9966 ------- estimated by Yves Gallot --------- 6000 5970 1,6498 7000 5001 1,3820 8000 4298 1,1877 9000 3765 1,0404 10000 3347 0,9249

As you can see, current (at 5000T) sieve efficiency for given parameters is about 2.0, i.e. sieving removes candidates two times faster than it's possible to do with LLR. We can run sieving for a while.

How long will CW sieving project lasts?

After completion of every 1000P range, I'll calculate number of factors removed in this range and corresponding sieve efficiency. When efficiency will be very close to or drops below 1.0, a 30-days official warning will be issued. Work will be available at least for 30 days. We can extend project a bit more to finish it at some round value, but amount of such extension will be unpredictable, so don't rely much on it.

Looking how efficiency drops, it's possible that shutdown sequence will be initiated after 7000P or 8000P.

My question would be why stop even when the efficiency is nearly 1:1?

If the September and December challenges can be switched, we could artificially add around 10 days to the life of the sieve.

My question would be why stop even when the efficiency is nearly 1:1?

- As already mentioned above, a goal of PG is to find primes. You cannot find primes in sieving. As for composites, if CPU can eliminate a candidate faster, do it on CPU.
- GPU is a powerful and rare resource (comparing to fleet of desktop PC's) which should used in most optimal way. For example, to find GFN primes or do other discoveries in projects which are optimized for GPU.

At current participation rate, we can do 1000P in 25 days. Considering Yves' prediction, it leads to following tentative plan:

- A 30-days warning can be issued at 8000P.
- After 30 days, the leading edge should be slightly above 9000P. Final stop point will be selected at 9500P or 10000P, giving efficiency ratio of 1.0 or slightly below.

Of course these numbers can be changed depending on real participation rate. But the main rule is: the project should not go above 10000P.

Did you mean 10000T instead of 10000P?

Nope, 10 000P is correct.

Anyway, the leading edge is currently a bit above 9000T, so I wonder whether we'll be getting a 30-day warning soon. I'm going to upgrade my CW-Sieve badge while I still can :)

Are you in the discord? I post bi-weekly updates on Mon and Thursdays.

But since you asked, current plan is to end at 9500P. Production is down in the dumps right now due to GFN challenge sucking up all the GPU hardware, which is putting the ETA at Jan 18th; before that, it was hovering around the 4-9th mark.

As you can see, current (at 5000T) sieve efficiency for given parameters is about 2.0, i.e. sieving removes candidates two times faster than it's possible to do with LLR.