Banking local memory causes nondeterminism

A central part of my design is a local array declared as follows: float arr[1 << 16]; This works well, but to increase throughput I'm trying to bank it as follows:

float arr[1 << 16] __attribute__((numbanks(8), bankwidth(4)));

So 8 banks with every float element interleaved. This works fine on the emulator, but causes nondeterministic when I synthesize the design. What could be the reason for this? I'm using Quartus 21.2 and an Agilex 7 board with product code B2E2_8GBx4. If I'm doing something wrong can you please suggest a workaround?