It can be pretty difficult to find an AMD-based small form factor (SFF) kit that is compatible with recent ESXi releases. The large majority of these AMD SFF kits all use Realtek for their onboard network adaptor and Realtek does not have a driver for ESXi, nor have they expressed any interests in creating one.
This is also quite unfortunate, as there have been some pretty interesting AMD kits that have been released in the market, but they can not be fully utilized due to the network driver challenges. Prior to the holiday break, fellow VMware colleague Hilko Lantinga, who works in our End User Computing (EUC) division had reached out about a new AMD kit that he had just acquired and has been very happy with after putting it through its paces, especially from a power efficiency standpoint for his updated VMware Homelab.
The system is from a vendor called ChangWang, which I was not familiar with and the AMD kit is the CW56-58, which is based on an AMD Zen 3 architecture (7nm) and what stands out immediately is that this system includes four built-in network adaptors (hint: they are all compatible with latest ESXi 8.0 release)! I asked if he would be willing to sharing his thorough review with the community as this could be a really interesting system and below is Hilko's detailed write-up in his own words.
With the current EU energy crisis (plus the upcoming CO2 tax) and to remove some heat and noise, I’ve been thinking about getting a small power efficient device that can run the eight VMs I always want to be on. I’ve got two big dual AMD EPYC Milan hosts with loads of NVMe disks and GPUs and a SMB 10GbE switch. Ideally the system would have 8-16 Zen3 cores (so I can vMotion), 64-128GB of RAM and 2-4 2.5-10GbE ESXi 8 compatible NICs.
When I saw the CW56-58, my interest was sparked, could this be my ultimate noise/heat/power saver that would put my large hosts in standby 90% of the time?
The CW56-68 can be purchased on AliExpress from Topton, who is a reseller they ship to both US/EMEA.
The CW56-58 comes with two AMD Cezanne U (Zen 3, Ryzen 5000 Mobile) and a Barcelo refresh CPU options:
|CPU||AMD Ryzen™ 5 5600U||AMD Ryzen™ 7 5800U||AMD Ryzen™ 7 5825U|
I opted for the Barcelo refresh 5825U as the price difference was slim. What is quite interesting about these CPUs is that they consist of a single Core Chiplet Die (CCD), so for the 5825U all eight cores share a single L3 cache. This means that when you create an eight vCPU VM all the threads can work together instead of having two groups of four.
sched-stats -t ncpus
1 NUMA nodes
Note: Recommended VM setting would be to have the cores per socket equal to the number of vCPUs. The CPUs have 16 usable lanes of PCIe 3.0.
The case feels a lot sturdier than pictures show, it feels quite luxurious, everything is thick and heavy with all-aluminum, almost like the better fanless cases as if every side is a heatsink. The case is 16.8cm (6.6”) wide, 13.17cm (5.2”) deep and 6.3cm (2.5”) heigh including the 3mm (0.12”) feet.
The front of the case has the power button, CMOS reset pinhole, one full-featured Type-C USB interface with data/display/reverse charging, two Type A USB 2.0 and two Type-A USB 3.1 ports as well as a Realtek RT5170 SD card reader. It’s a bit weird that the native USB 3.2 Gen 2 from the CPU doesn’t seem to be used, the Type-A ports might be coming from the chipset. The latest BIOS even has a UEFI module for booting from the card reader. While I don’t recommend using a SD card for booting ESXi, there is nothing wrong with installing from SD card.
The top of the case has an intake to what ChangWang calls a coppertube turbine, a blower style fan. The CPU is on the back of the motherboard, and it is mounted upside down in the case.
The rear of the unit has an HDMI and DisplayPort as well as 4 x RJ45 Intel I226-V 2.5GbE ports, a DC input and a screw for a ground lead.
Talking about cooling, I strongly recommend replacing the thermal paste of the blower style cooler. ChangWang used Xinyue 7921, but it was absolutely rubbish even when new. Swapping with Noctua NT-H1 (would recommend NT-H2, but had none left) made a 12 degrees C difference, and over time the difference would probably become larger and larger. You can remove the cooler by removing the four black screws on the motherboard, two between the M.2 drives and two under the SODIMMs, the four metal-colored screws can be used to remove the motherboard. For extra cooling it would be possible to put 1mm thick thermal stickers between the blowerfan and top of the case.
The BIOS allows you to choose (Advanced->AMD CBS->NBIO->SMU->System Configuration) between 10/15(default)/25/35 Watt operation, apparently older BIOS versions even had 45W and 54W, but that was removed.
I can imagine why, with the default thermal paste even 35W operation isn’t sustainable, I had to cancel my test when the CPU hit 99 degrees C and was still climbing rapidly, as I did not want to test what happens at the maximum operating temperature of 105 degrees C of the CPU.
With the new thermal paste all modes worked fine, even for an hour of CPU torture.
|BIOS Setting||10W||15W (def)||25W||35W|
|Short CPU Power Limit||18W||30W||48W||60W|
|Total System Power Short||39W||55W||84W||120W|
|All Core Turbo Short||2.7GHz||3.1GHz||3.8GHz||4GHz|
|Short Temp||70° C||72° C||76° C||83° C|
|Long CPU Power Limit||15W||25W||37.5W||42W|
|Total System Power Long||33W||35W||55W||80W|
|All Core Turbo Long||2.4GHz||2.4GHz||3.1GHz||3.7GHz|
|Long Temp||64° C||64° C||67° C||77° C|
|W11 Idle Power||10W||10W||10W||10W|
I decided to go for 25 Watt, this gives me a sustained All Core Turbo around the same speed as my dual AMD EPYC 7713 (3.1GHz), 5 minutes short All Core Turbo of 3.8GHz, single core Turbo of 4.5GHz and only 55W during long powerful operations, all while maintaining the same low idle power usage and temperatures I’m comfortable with. It might be more energy efficient during intended operation as 15W as the machine would be able to go back to idle quicker. I would not recommend the 10W option as there hardly are any power savings, nor would I recommend the 35W option, with the default thermal paste it doesn’t work well, plus the peak power usage is above the 90W rating of the (probably not that great) included power supply.
The unit is silent when idle, around 45dB when the fan really kicks in and steadies around 40dB, both measured at 50cm distance. Now while this might sound as a reasonably quiet system, it isn’t. The noise has a high pitch to it, which makes it appear a bit louder than it is. Fine for in a fuse/cupboard, but not something I’d like to have on my desk in the office.
While most Chinese manufacturers don’t provide BIOS files, ChangWang actually does, which is good to see. Note that there are two versions of the BIOS, there is one with the reversed order of the NICs. My system came with the latest 0.22 BIOS from the 9th of October 2022 and runs microcode 0x0a50000d, which seems to be the latest AGESA V2 PI 184.108.40.206.
I used two Crucial CT32G4SFD832A 32GB DDR4-3200 CL22 DIMMs for a total of 64GB, the maximum of this board. By default, the board sets 1GB of memory aside for graphics. As the ESXi console doesn’t need it and my workload does I changed that to 64MB in the BIOS (Advanced->AMD CBS->NBIO->SMU->GFX Configuration->UMA Frame buffer size). I did a couple of rounds of memtest x86 before installing ESXi, they passed.
The M.2 slots only have screw holes for 2280 drives, so make sure you don’t buy 110 or 60mm drives but get 80mm ones. There was a Black Friday deal on the Samsung 980 Pro 2TB, I opted for two with heatsink as in systems this small NVMe controllers can easily overheat. As you can see in the image it almost doesn’t fit, the top drive is ever so slightly angled around the unused ATX power input. Any heatsink wider or higher, even by a hair, will not fit. The 980 Pro is PCIe 4.0, while the system is 3.0, but PCIe is downwards compatible, and the 4.0 speeds might become useful if upgrading to another system later.
Installing ESXi 8.0 is easy as everything in the system is compatible. Secure Boot just works and even for the NICs you no longer need the Community Networking Driver Fling.
The idle usage without VMs in ESXi is 15 Watts, with a single idle VM it is 16 Watt. The sustained boost frequency is 3375MHz with a power usage of 55 Watts and the short boost frequency is 3900MHz with an average power usage of 84 Watts.
What does not work is TPM, because fTPM exposes a CRB interface and not TIS/FIFO as required by ESXi. Disabling TPM in the BIOS (Advanced->Trusted Computing->Security Device Support->Disabled) clears the warnings and error messages. You can still use vTPM without a physical TPM device.
What also doesn’t work is vMotion from AMD EPYC Milan without EVC, as Process-context identifiers (PCID) aren’t available, which is weird as it should be a Zen 3 CPU feature. Either it isn’t available on mobile CPUs or maybe something isn’t correct in the firmware. But with EVC set to Zen 2 (AMD Rome didn’t support PCID) on both hosts vMotion from/to AMD Milan works fine.
Both my Milan hosts and the 10GbE switch idle around 700 Watts. Even if I put one host in standby, we are still talking about 400 Watts. With both hosts in standby, the 10GbESMB switch powered off and a consumer 10GbE switch running, the usage is 45 Watts. The average over 24 hours of this unit with 8 mostly idling VMs including vCenter, Active Directory, SQL Server, VMware Horizon, OPNSense, etc. (average total usage is 970MHz) is just 21W! And when I need to benchmark something on customer like equipment or test something with GPUs I can power on or put hosts in standby straight from vCenter, while for minor tasks I can use the CW56-58.
Total power savings per year would be 2600kWh and with the current energy prices, it only takes a couple months to get a full return on my $1000 investment. I could save another 250kWH by powering off IPMI too. Needless to say: I’m keeping it.