A SBC for computing - your thoughts

[Werner]

Imagine a new SBC is developed specifically designed for computing tasks and clustering but should also be available for a rather low price.

Which features should it have or not have? What do you think? Is it even possible?

[arox]

Manufacturers will always add all hardware features that does not cost a lot : i.e. crappy cheap wifi and so on, because they do not bother to provide good drivers.

 

In the other hand, RAM is always priced a lot and thermal design bad. If I want a SBC for specific use case like computing (and not a X86 server), I want it to be passively cooled and energy efficient. Storage is also a choice : either emmc, either a good bus design for drives.

 

I simply cannot find a SBC for a use case because they are all designed for "general purpose" like a RPI or expensive with no guarantee of software support.

 

[gounthar]

No WiFi, no SDCard, M.2 socket, eMMC, SPI Flash, USB-C (power plus host), 4GB RAM, big low profile heatsink, RTC, and GPIO with I2C, SPI, and pins with USB. As small and low height as possible.
Options would be NPU, more RAM, CSI and of course a SOM instead of the whole SBC.

My $0.02.

[arox]

1 hour ago, gounthar said:

big low profile heatsink

 

Many seems to think that theirs has to be very long and powerful (if you understand what I mean). In fact, a well designed enclosure can assume this function. And in any case the quality of the thermal interface is more important than anything else.

[gounthar]

The thing is I want to be able to incorporate my SBCs in some 19" rack (1 U preferably), in desk grommet, or any other type of enclosure, so I don't want a well designed enclosure, as it won't help me.

[lanefu]

I want a stripped down board that has real 802.3af POE

[SteeMan]

When designing a good sbc for general compute/server tasks I think you need a set of features that enable both a 'desktop' as well as 'server rack' deployment.  The reason for this is the evaluation process someone will likely undertake in order to buy into the boards features. No one is going to buy 32 boards for a 'server rack' deployment as the first purchase.  Instead they are likely to purchase one or a few to evaluate first.  That evaluation is not going to happen in a rack mount, but instead will happen on a desktop.  Once someone is comfortable that the base board works for their basic needs (i.e. the software and general hardware works), then they will explore the 'server rack' deployment options as they plan to scale a use of the board.

In my opinion therefore you need to make sure you have the features necessary to have a good evaluation experience on the desktop for the board ultimately to be successfully purchased in larger quantities for server work.  One example of this is an hdmi port.  While an hdmi port is completely useless in a server deployment, it can be quite useful during board evaluation on a desktop.  Another example is cooling as mentioned in the above posts.  I think you need to have good thermal design for both deployment scenarios (both as a desktop board and in a server rack mount), which might require different heat dissipation strategies for the different environments. Finally POE while likely unnecessary for a desktop evaluation is critical for a server deployment.

 

My ideal feature list would be:

1gbit POE ethernet port

4GB ram

32GB emmc (or more optional)

good external storage options (m.2 or other)

hdmi port

2 or more usb ports (at least one being usb3)

power port for non POE usage

optional case for desktop use with good thermals

optional rack mount with good thermals

 

The two things I think it shouldn't have:

- no wifi/bluetooth

The reason I say these are not desired is that good wireless (good antenna's, good software support) is difficult to design into a board, it isn't needed in the server rack case and can be accomplished better with a usb addon for the desktop case without incurring the added cost to the base board.

- no SD card

The reason I wouldn't include an sd card slot is if emmc is standard, that will be the preferred deployment storage media.  You only need another option to install/update the internal emmc and usb should be sufficient for that.  The sd support then just becomes an added cost with no real long term need.  It does require that booting from usb be well supported by the firmware.

 

Such a board would span a lot of use cases from general purpose single desktop use case to hundreds of boards deployed in dense rack configurations.

 

My personal experience is that I try things out first by evaluating one of something, then scale up to a few, and ultimately more as each step of the evaluation process shows the product is capable of the next deployment step.

 

Finally I'll mention price.  In my opinion you likely need the above described board at a price point no more than a RaspPi.  Given the large ecosystem and mind share built around that platform, and it is already capable of doing the above (although not well in many respects), you can't have something like this be at a 'high end' premium price point and expect it to be successful.  Price will to an extent drive the evaluation process.  If the price is considered too high, then people won't even start the evaluating, they will just stick with what the everyone else uses.

 

[lanefu]

IMHO I'd omit HDMI or any sort of output other than console.    If the eMMC is a detachable module, then yeah eliminating SDCard is a good idea.

[arox]

2 hours ago, SteeMan said:

you likely need the above described board at a price point no more than a RaspPi.

 

With 32 GB emmc ?!?

 

That is an option that generally costs the price of a RPI.

 

[NicoD]

I just want 32 N1 cores. No need for any other I/O than a Gbit ethernet port. 
More would be nice. My M4V2/N2+ do everything I need. Just for rendering and building I can use more performance. Being able to do these jobs on a dedicated device would be nice.
Tho a workstation with 32 N1 cores is more awesome.