The next team challenge at PrimeGrid will be on March 14 – March 24, 12:00 UTC, at the "Seventeen or Bust" subproject.
This is a CPU-only subproject which is checking very large candidates for primality: Currently, its candidates have 10 million digits. Only the PrimeGrid subprojects GFN-21, GFN-22, and GFN-DYFL ( = GFN-22 at a higher search range) look at larger candidates, and these are GPU subprojects. (OK, GFN-21 also has a CPU application, FWIW.)
I looked up the
current search range of Seventeen or Bust and checked its FFT sizes on a Haswell CPU. Haswell and all later Intel CPUs which don't implement AVX-512, and several AMD CPU architectures including Zen 2 and certainly Zen 3, will use the same FFT implementation here, which is "all-complex FMA3 FFT".
SoB-LLR candidates | current FMA3 FFT lengths | last-level cache size demand |
---|
21181*2^n+1 | 2880K … 3M | 22.5 MB … 24 MB |
22699*2^n+1 | 2880K … 3M | 22.5 MB … 24 MB |
24737*2^n+1 | 2880K … 3M | 22.5 MB … 24 MB |
55459*2^n+1 | 3M … 3200K | 24 MB … 25 MB |
67607*2^n+1 | 3M … 3200K | 24 MB … 25 MB |
Zen 2 will perform poorly at this, because its level-3 cache is divided into 16 MB segments. (In APUs, it's much less.) Zen 2 will therefore have to perform a lot of main memory accesses and will consequently be unable to keep its FMA units busy.
Zen 3 has got 32 MB level-3 cache segments, one per each compute die. (Zen 3 based APUs will have less.) The Zen 3 CPUs should therefore perform very well in the SoB-LLR subproject, provided that there is exactly one LLR program instance running per each compute die. (That is, one SoB task at a time on 5600X and 5800X, and two tasks at a time on 5900X and 5950X. Thread count per task should be 6 or maybe 12 on 5600X and 5900X, and 8 or maybe 16 on 5900X and 5950X.)
My old 22-core Broadwell-EP CPUs have 55 MB unified L3-cache and will therefore perform OK with 2 simultaneous SoB tasks per socket.