That doesn't make sense to me. Memory bandwidth refers to the throughput of moving data from memory. Even if FP4 calculations aren't natively supported, you still move 4 bits of data per number from memory and then cast it to a FP8 or FP16 or other higher-precision number, right?
Llama.cpp already moves 1.5bit to the cores, they are converted in SRAM which has much higher bandwidth.