By the way, here are the results of naive tests (no JMH, so be careful with conclusions) on GraalVM11:

Boxing addition completed in 22157 millis

Specialized addition completed in 1840 millis

Nd4j specialized addition completed in 1309 millis

Viktor addition completed in 1966 millis

Parallel stream addition completed in 1457 millis

Automatic field addition completed in 1773 millis

Lazy addition completed in 14157 millis

ND4J uses OpenBlas under the hood. And I think @Iaroslav Postovalov told me that is uses parallel execution. I wonder if there is a large overhead on top of BLAS. Because the results are very close.

Nd4j uses MIMD

SIMD is not parallel. Viketor and JDK11 also use SIMD

I said MIMD

I did not enable it for this test. And actually it did nothing when I've uncommented your lines in build