Warning about Epyc Naples
Warning about Epyc Naples
This approach aligns closely with a dual CPU board setup, where memory bandwidth is shared effectively. For virtual machines with eight or fewer vCPUs, selecting the right cores should work fine—up to four physical cores per die should suffice. The system would manage memory channel assignment automatically when you choose the appropriate number of cores. The ROME6U-2L2T model has only six channels, which adds another layer of complexity.
Each memory area is controlled by a NUMA zone/node. NUMA refers to memory rather than CPU threads—it just coincides in x86 servers since cores are tied to these zones. The count of cores per zone can vary from memory size. When you place a VM's CPU threads onto a core, it automatically links to the same zone because each core must belong to one. The hypervisor keeps track of which cores go where, choosing local ones for efficiency within a VM. But this isn't the only rule: if cores are overwhelmed, some may move to another zone. Another case arises when RAM allocation is excessive—for example, giving 65GB to a single VM with 128GB total split across sockets. Memory would be split between nodes; remote accesses would use QPI links between zones for any needed data. The main point remains: NUMA focuses on memory placement, not CPU count, so understand per zone limits before scaling VMs. This advice mainly fits EPYC 1st Gen and Intel multi-socket setups. Multi-socket models can span smaller zones within one socket, while Intel systems don't support this across sockets. A NUMA diagram for a dual 7713 server shows how cores are distributed across nodes, helping you know the maximum memory per VM.
Imagine you configure it with ZFS V2.0 on FreeBSD 13.0 and feed it twenty SSDs. So far things feel odd. Data moves through the CPU at full capacity. On Linux I didn’t replicate that performance. If this is intentional, I’d display it proudly. Others have warned HPE about questionable server practices—no BIOS updates without an account or plan, which is a big concern. I’d likely install it on one of my servers right away. If memory problems disappear, I might even use it as my primary hypervisor host. Adding 0.5TB of RAM would mean only one server handles everything.
Access is completely open, but a service pack account is required for the ProLiant (SPP) ISO. All firmware and drivers are included in one package. You also need an account for the System ROM firmware, though updates are usually handled through other methods if you prefer. It’s not impossible to obtain these resources.
I go with the usual standard platform choice. ASRock Rack & Supermicro are on my list too. Talk about. The Xeon Platinum 8176 looks promising—two units could definitely boost a VM server setup.
Have you explored Truenas Core (Freenas) for potential fixes? You’re considering Truenas Scale paired with HDDs and an all-SSD pool for VMs and Docker. I’ve purchased a 7401p for $105, which isn’t tied to any vendor (the seller recently upgraded his own system). I won’t miss it and will switch to Rome or Milan when prices fall. I’m planning to use a ROMED6U-2L2T board that’s confirmed to work with Naples via v1.0 BIOS, so it’ll be ready once you decide to upgrade.
I chose FreeBSD 13.0 for its strong command-line management features. I previously ran FreeNAS on TrueNAS and it performed well, but I prefer engaging with the FreeBSD community to learn optimization techniques. At $105 for a 24-core system, it’s a solid investment. The Naples performance claims hold true, though the decision ultimately depends on your goals and budget. Naples offers affordable high I/O capacity. With 64 PCI_e lanes, it’s ideal for parallel workloads that can be split across multiple chips and properly configured RAM slots. Since CPU constraints are a factor, you can upgrade to a ROME chip later when prices drop.