If your 550MB/s 4KB@QD32 is from CDM, and you're running on stock settings, you are likely CPU-bound by one core/thread. Try AS SSD (wich uses 64 threads), or an IOmeter config with # of workers = threads (8 or 16?), and scale QD.
I'd be happy to assist you with making a couple of graphs of IOmeter data to illustrate the scaling. I've got data for Computurds 8R0 barefoot array with fastpath, so we could compare them