AMD Ryzen Powered
On the AMD Ryzen side, there are two motherboards that are possible. We have both the ASRock Rack X470D4U and X470D4U2-2T in the lab, and Will has reviews coming.
While many may prefer the X470D4U2-2T with dual 10Gbase-T, I personally would build around the X470D4U here instead. We have so many inexpensive 25GbE NICs, and I prefer SFP+/ SFP28 using SFP+ to 10Gbase-T adapters when necessary. It is actually less expensive to get a 25GbE card in the X470D4U than it is to get dual onboard 10Gbase-T in the X470D4U2-2T. This was a top-2 finalist but we decided against it.
AMD EPYC 3000 Series Powered
The AMD EPYC 3000 series in a mATX, specifically an AMD EPYC 3351 12-core platform with remote management and the I/O features exposed, would have been just about perfect. Here is the Wallaby development board, which is very close to what I would want.
At the time we are doing this piece, the currently shipping platforms are mostly industrial-focused or mITX. Most are single-die solutions and do not offer 10GbE onboard.
For a 4-bay solution this may be absolutely perfect, but for what we are aiming for we wanted a mATX platform.
AMD EPYC 7000 Series
The AMD EPYC 7000 series could have been very interesting, but it presented almost exactly the opposite set of problems that we had on the Xeon Scalable line.
On one hand, cooling the CPU is very easy even though it is a higher-TDP platform. There are plenty of Threadripper closed-loop liquid cooling solutions that would have cooled the CPU in our SilverStone CS381 without issue. Since the physical mounting points of the Threadripper and EPYC 7000 series platforms are the same, this was easy and we had the liquid cooler on-hand.
The harder part was the motherboard. mATX is simply too small of a form factor for the EPYC 7000 series. With eight memory channels and a large socket, it takes up a lot of PCB space. Further, one of the most significant benefits would be getting a lot of PCIe I/O, but that would be mitigated on mATX due to form factor’s physical size. Great platform, but one needs a larger motherboard for it to really make sense. We did, however, Introduce the Ultra EPYC AMD Powered Sun Ultra 24 Workstation, which was a great use of that platform even in a smaller chassis.
What we used: Xeon D?
It feels a bit strange saying this in 2020, but we ended up using a Xeon D platform for this build. Using a 45W TDP SoC with 10GbE was extremely attractive. We ended up with the ASRock Rack D1541D4U-2O8R as our platform of choice.
Frankly, this is still too much CPU for a FreeNAS or TrueNAS core build, so we probably would be better off with a Proxmox VE solution here. With ZFS and Linux becoming cozier, that would have made sense. The other nice thing is that the onboard SAS3 (Broadcom/ LSI SAS3008) uses a SFF-8643 connector, as does the backplane on the CS381. It just so happened, we had those cables on hand.
Here, not only is the TDP of the CPU lower than the Xeon E-2200 and Ryzen solutions, but it is also lower when one looks at the 10GbE networking TDP and chipset TDP. Also, the SAS3008 controller is effectively a universally supported solution these days, whereas some OSes are not as well suited to run on chipsets. The lower power consumption aligning with the decision that we wanted a lower-core count and memory solution meant that we went back to Xeon D.
Taking a moment here, that is borderline absurd. Five years after the Xeon D-1540 release, we are still going through a process and the Xeon D-1541, the updated part with the SR-IOV fix is what we are using. Then again, it shows the power of embedded parts and why we need products like the EPYC 3000 series for disruption.
Next we are going to look at some of the other build bits in this system before getting TrueNAS Core installed.