Would love to get feedback from the members here on possible issues, as well as would you go rdimm/lrdimm… (use case is devops/ml and db over proxmox virtualization)
QVL is basically just memory the board manufacturer has explicitly tested. For the pedantic notes:
Crucial’s retail shop doesn’t specify which die revision you’re getting (the “E1xx” tail of the model), so it may not match what ASRock Rack tested.
The /BCM variant is a different revision of the board, not just component configuration, and doesn’t have any QVL entries yet. (For a comparison point, the ROMED8-NL is a configuration of the ROMED8-2T and has the same QVL list.)
All that said, the QVL list is certainly a good place to start from, and avoids potential hassles. You’d want to follow it if building many systems as an OEM or integrator, since it avoids you having to do more component validation yourself.
As for RDIMM vs LRDIMM, the short answer is: you don’t need LRDIMMs here.
LRDIMMs have additional buffering on the interface to reduce electrical load on the memory controller, at the cost of a little additional latency and therefore reduced peak bandwidth. They’re designed for high density configurations where there’s a need for more ranks of memory in a single channel.
This board supports 1 DIMM per channel, and you’re looking at 2-rank 64GB RDIMM modules, which is the max for them and is perfectly fine. If you needed 128GB or more per module, you’d look at LRDIMMs and 3DS modules.
On a board supporting 2 DIMMs per channel (16 slots per CPU socket), you’d be deciding between lower clocks with 16 RDIMMs, or higher latency with 8 LRDIMMs: see page 8 of the memory population guidelines.
Neither is an concern for you unless you decide to start with 1TB for this build
