I implemented the algorithm in Appendix C.1 and successfully recomputed table B.1, as well as it's equivalents in FIPS 186-4 (notice that the question's rendition of table B.1 has the probability wrong for the right column, butlast and last box, where that probability should be $2^{-128}$ and $2^{-144}$ respectively).
The values for 3072 and 4096-bit keys are those in the second and third rows of cells of table B.1. Here is an extension to common key sizes, also covering obsolete ones.
Minimum number of rounds of M-R testing for RSA key generation, assuming trusted random $p_1$, $p_2$, $q_1$, $q_2$, $p$ and $q$, from FIPS 186-5 table B.1 unless otherwise noted
| Parameters |
Error probability $2^{-100}$ |
Progressive error probability |
1024-bit public modulus1,2
$p_1$, $p_2$, $q_1$ and $q_2$ > 100 bits
$p$ and $q$: 512 bits |
For $p_1$, $p_2$, $q_1$ and $q_2$: 38
For $p$ and $q$: 7 |
Error probability $2^{-80}$
For $p_1$, $p_2$, $q_1$ and $q_2$: 28
For $p$ and $q$: 5 |
1536-bit public key2,3
$p_1$, $p_2$, $q_1$ and $q_2$ > 120 bits
$p$ and $q$: 768 bits |
For $p_1$, $p_2$, $q_1$ and $q_2$: 35
For $p$ and $q$: 5 |
Error probability $2^{-96}$
For $p_1$, $p_2$, $q_1$ and $q_2$: 33
For $p$ and $q$: 5 |
2048-bit public modulus
$p_1$, $p_2$, $q_1$ and $q_2$ > 140 bits
$p$ and $q$: 1024 bits |
For $p_1$, $p_2$, $q_1$ and $q_2$: 32
For $p$ and $q$: 4 |
Error probability $2^{-112}$
For $p_1$, $p_2$, $q_1$ and $q_2$: 38
For $p$ and $q$: 5 |
3072-bit public modulus
$p_1$, $p_2$, $q_1$ and $q_2$ > 170 bits
$p$ and $q$: 1536 bits |
For $p_1$, $p_2$, $q_1$ and $q_2$: 27
For $p$ and $q$: 3 |
Error probability $2^{-128}$
For $p_1$, $p_2$, $q_1$ and $q_2$: 41
For $p$ and $q$: 4 |
4096-bit public modulus
$p_1$, $p_2$, $q_1$ and $q_2$ > 200 bits
$p$ and $q$: 2048 bits |
For $p_1$, $p_2$, $q_1$ and $q_2$: 22
For $p$ and $q$: 2 |
Error probability $2^{-144}$
For $p_1$, $p_2$, $q_1$ and $q_2$: 44
For $p$ and $q$: 4 |
6144-bit public modulus3
$p_1$, $p_2$, $q_1$ and $q_2$ > 230 bits4
$p$ and $q$: 3072 bits |
For $p_1$, $p_2$, $q_1$ and $q_2$: 18
For $p$ and $q$: 2 |
Error probability $2^{-160}$
For $p_1$, $p_2$, $q_1$ and $q_2$: 47
For $p$ and $q$: 3 |
8192-bit public modulus3
$p_1$, $p_2$, $q_1$ and $q_2$ > 260 bits4
$p$ and $q$: 4096 bits |
For $p_1$, $p_2$, $q_1$ and $q_2$: 16
For $p$ and $q$: 1‡ |
Error probability $2^{-176}$
For $p_1$, $p_2$, $q_1$ and $q_2$: 50
For $p$ and $q$: 3 |
1 From FIPS 186-4.
2 Obsolete.
3 Extrapolated, using FIPS 186-5 appendix C.1 for the number of rounds.
4 This bound is arbitrary. Further, absent any certification requirement, the only rationale for $p_1$, $p_2$, $q_1$ and $q_2$ this large seems to be reuse of code for the generation of $p$ and $q$ that requires them. While the use of $p_1$, $p_2$, $q_1$ and $q_2$ could be justifiable when generating $2^{40}$ 1024-bit RSA public keys and requiring infinitesimal probability that any succumbs to a factorization effort concentrating on applying Pollard's $p-1$ or Williams's $p+1$ to many public moduli rather than ECM, or GNFS to one modulus, that rationale fails for larger public keys.
‡ But perhaps make that 2.
