Also are there any plans for COM-HPC stuff?
Image for reference:
COM-HPC is pretty cool stuff:
Oh, whoops thought this was the only Standard 
- COM Express: COM Express - Wikipedia
- SMARC: Smart Mobility Architecture - Wikipedia
- QSeven: Qseven - Wikipedia
- OSM: https://www.adlinktech.com/en/Computer_on_Modules_OSM
- ETX: ETX | COM | ADLINK
Image: XKCD 927
2 Likes
My understanding is that everything is soldered to the board for bandwidth and speed reasons.
I agree though, if it doesn’t hinder performance, I would be interested in seeing this explored in the future.
No it is not. I it is a Mini-ITX board, but with a limited PCIe4.0 slot (only 4 lanes, and I think 25W), given limitations of the Strix-Halo regarding available PCIe lanes. There are no bottom side connectors like for those computer on module systems and that is simply also not the use case Framework is after. I don’t know how these connectors work, are they based on PCIe or something entirely different? Chances are Strix Halo as a platform doesn’t support it.
Can you explain why this would be valuable?
-
It’s a Mini ITX board which is a standard board type
-
Because it is Mini ITX, you’d have two boards you are paying for that one isn’t much bigger than the other
-
Swapping out the “CPU/RAM" board would only make upgrades only mildly easier than just replacing the entire board
I mention all of this, because it makes a lot of sense from like a Raspberry pi/single board world (see Ri Pi3,4,5 and their CM versions) because there is no “standard” when it comes to Single Board computers and those a relatively low power devices that tend to get crammed into extremely tight spaces. These are fully fledged computer boards pulling 100+ W that fit into relatively large cases…. With a crap ton of high speed PCIE lanes (vs single board computers with their 1-4 PCIE 3.0 lanes)
Oh, I posted this cause I thought the main concern with non swapable RAM/CPU was bandwidth between the CPU and RAM not so much the IO. Having COM-HPC would “solve” that problem.
As to why I think that would be valueable:
- You can Choose you IO
- Wanna Build a home server with multiple disks?
- Go for a carrierboard with more SATA
- Wanna build a crazy cluster?
- Go for a board with high speed interconnect.
- Wanna build a server for hosting a website?
- Go for a board with an integrated IPMI Module
- Wanna Build a home server with multiple disks?
- You can swap between diffrent architectures (test out fancy RISC-V chips)
- You can choose between a system with unified memory and one with seperate CPU/GPU units
So in summary, you can for example buy it as a desktop and choose later what to do with the old compute module. IO determines the most suitable application, more than the compute.
Yes, it will probably cost more than regular ITX boards. But I’m not so sure if it will be significantly more expensive. ATX and ITX boards are not so dissimilar in price, considering available PCB area.
I think those are worthy objectives. I don’t have much specific knowledge of Strict Halo, but if I understand correctly what others have posted here, the chip itself has I/O limitations that will essentially propagate to the I/O board, resulting in little benefit for the effort expended in a COMP-HPC implementation.
Yeah, having limited PCIE lanes probably makes this less optimal. But with custom IO you can at least choose what those limited lanes are used for.
But again, I am not sure if the COM-HPC connector introduces serious limitations. I would need a more foundational understanding of the standard and PC hardware.
1 Like
But you don’t need a separate I/O board for that. 8 of those 16 available lanes are M.2, 4 more are a PCIe port. The remaining 4 lanes are reserved for Wifi and LAN. So just get an M.2 or PCIe adapter.
Or a M.2 Oculink adapter. The image above suggests that you might be able to omit a tile on the front or use a custom printed one, and lead out the oculink cable sanely out of the enclosure.