Pictured above is the brand new DeskPi Super4C put in in an 8U mini rack. The Super4C is a 4-node Raspberry Pi CM5 cluster board that solves two ache factors I had with the older Super6C.
I used to be testing this board across the similar time I helped kick off the SBCC 2026, the Single Board Cluster Competitors for college kids. A dozen or so college groups squared off to run one of the best mini HPC cluster with a price range of $6,000, and a pair days to benchmark six HPC workloads.
I made a decision to populate the Super4C DeskPi despatched me to check with 4 16GB CM5s I acquired final yr, again once they have been $125 every. They’re about $300 at present, after the newest spherical of value will increase.
I made a video reviewing the Super4C and a brand new mini rack from Waveshare, the HomeRack, and I posted it on YouTube:
In that rack construct, I additionally examined Waveshare’s HomeRack Pi 7″ Contact Panel mount, in addition to a Tripp-Lite mini UPS that slots properly into the underside. Should you’re within the {hardware}, go watch the video.
The Goldilocks Cluster
The primary factor I wished to level out on this weblog submit are the 2 annoyances DeskPi mounted within the Super4C—which I feel they mounted by lowering the variety of Compute Modules from 6 to 4 (thus gaining precious board house):
Distant Administration: There’s now an ESP32 you possibly can entry over WiFi (or, quickly, Ethernet) for ‘lights out’ administration of the cluster—no less than for energy management and monitoring.Redundant Ethernet and Energy: This board has two Ethernet connections to every CM5 slot (one 2.5 Gbps NIC off the USB 3.0 bus, and one 1 Gbps connection by the CM5’s personal Broadcom NIC). It additionally has twin 19V DC barrel jack energy inputs, and can change between them to whichever one has the upper voltage.
The video has extra element about different board options, however I point out that it is sort of a ‘Goldilocks’ board for the SBC Cluster fanatic; it has simply sufficient options to resolve the ache factors I had with the Super6C, however not a lot complexity the value is pushed past a hobbyist’s price range.
I do not advocate a cluster like this for worth. The explanation you’d purchase one thing like that is for HPC tinkering, on a tiny scale, at your desk, with no bunch of power-hungry used mini PCs or outdated servers heating up your room.
It is also helpful in another area of interest use circumstances:
A multi-camera / multi-node imaginative and prescient processing system, with one or two cameras per CM5 (as much as 8 cameras complete, with an AI accelerator for every Pi)A tiny 4-computer board that has an HDMI show on every nodeA small construct farm for Arm tasks or CIA tiny redundant edge cluster for light-weight functions
Should you simply wish to construct slightly homelab, although, and you do not care about studying networking or multi-node deployment/coordination: purchase a mini PC. You will get an honest one these days for $500-600 with 32 GB of RAM, even with inflated reminiscence pricing.
As it’s, a minimal cluster construct with 4 Pi CM5s with eMMC and a pair of GB of RAM will price practically $600. And the cluster as pictured above is nearer to $1,400! Like I stated, it is not about worth, versus a single beefy machine.
The Competitors
It is enjoyable to see how this cluster operating on 4 CM5s with a complete of 16 A76 CPU cores and 64 GB of LPDDR5 RAM compares to the SBCC scholar clusters.
Here is the way it stacks up operating HPL:
Not a foul exhibiting, general! In my kickoff speech, I instructed the scholars I suspected the group with a cluster of Orange Pi 5 Max boards would have an excellent shot on the title—and operating HPL no less than, they have been second!
Taking a look at effectivity, my Super4C cluster was additionally in the midst of the pack:
The opposite cluster (PlumJuice) that is neck-in-neck can be operating on Raspberry Pi 5s (9 of them), so it is good to see that consistency in effectivity from cluster to cluster.
The outlier in each circumstances was a cluster of 4 Minisforum X1 Professional-370 mini PCs—with their boards ripped out of them in order that they have been extra within the spirit of ‘Single Board’ computer systems 🙂
Here is a take a look at all of the specs for the clusters within the high of the HPL outcomes:
The NTHU group’s AMD nodes have been underclocked for higher effectivity (AMD and Intel appear to push default clocks approach past what’s needed, I assume to one-up one another), and so they additionally compiled HPL with ROCm towards the iGPU for a little bit of a speedup, too.
You’ll be able to name that dishonest, since most individuals do not take a look at Minisforum PCs and suppose “SBC”, however I name it a artistic interpretation of the principles.
Apart from, the NTHU group did not win the general competitors. Group Kent Ridge did, with a cluster of 16 Orange Pi 5 Max’s.
I want the SBCC existed again once I was in faculty. Because it did not, I am glad I can no less than Cosplay as a Sysadmin and construct my very own little clusters to match at present.
These cluster tasks do not make sense anymore when you simply must run a homelab. However they’re fairly enjoyable if you wish to be taught HPC with a cluster that sits in your desk.
