Curious if anyone has tried this or can provide input. Our scratch is on an m.2 PCIe 3.0 but is shared by cryosparc (often running 2x processes on 2 GPU’s) and relion, where we are running 3D classification w/o alignments. Working with lots of binned particles of various datasets at once and trying to think of how to speed things up. Our motherboard has a u.2 connector into which we could put a “Four NVMe M.2 to U.2 Carrier Shuttle” (OWC U2 Shuttle) and put two or three m.2 in RAID0 config to use as scratch. Would this make sense? The u.2 port supports PCIe 3.0 x4 NVMe.
What I’ve been doing is using 2 SATA SSDs and configuring them as a software RAID0. You’ll automatically get a a big performance bump with m.2 over SATA, and an even bigger one with u.2. So, setting up 3 m.2 drives attached to a u.2 connector and configured as a RAID0 should give you maximal performance.