Attached is a simple Java test code that exposes the problem. The code
executes N time the method generateRandomString in sequence and then in
parallel (50 threads)

The code uses the SecureRandom API (secureRandom.nextBytes(destByte))

Command line is:
java -Djava.security.egd=file:/dev/./urandom TestSecureRandom 1000000

Here are the runtime results (lower the better) they get on a Red Hat x86
system vs. on a T4-2 (I did the test on an internal T4-2 system):

Number of Calls x86 Sequence x86 Parallel T4 Seq T4 Para
1000000 1035 ms 2026 ms 5975 ms 9015 ms
2000000 1966 ms 3020 ms 11191 ms 17016 ms
3000000 2935 ms 5008 ms 16215 ms 26017 ms

We are constantly at least 5x slower on the T4 than on the x86 system.