Comment by mdtusz | Hacker News Reader

Comment by mdtusz | Hacker News Reader

Parent Full thread

This may seem somewhat silly to most people but if it isn't vaporware and can actually meet demand and fulfill procurement needs, could be a solid choice. Choosing and sourcing SBC's for hardware is an absolute nightmare these days because of the thousands of options, horrible supply chains, and essentially nonexistent documentation or BSP's for the vast majority of them.

Raspberry Pi's are undeniably the easiest SBC to work with, and for a huge portion of industrial, defense, and aerospace products and devices are a near ideal choice except for their "hobbyist" level attributes.

A robust Pi4 clone that supports nvme drives, has no Bluetooth or Wifi by default, and has something like ISO9001 certification to check the boxes for big enterprise contracts would sell like hotcakes. It's somewhat surprising to me that the Pi foundation hasn't made something to this tune yet (the Pi CM4 may as well be vaporware - they're basically impossible to find anywhere even in single-digit quantities).

> supports nvme drives

Before that, I'd honestly like to see TPM 2.0 support baked in first.

> has no Bluetooth or Wifi by default

Alas, not the case here[1]. Genuinely looks like they just stuffed a consumer RPi CM4 on a custom carrier board wrapped in metal, pushed through the easier boilerplate environmental checks like all their other SBCs, then put out a press release.

Not sure how I feel about the interconnect choice; power via Omnetics proprietary (28 VDC input, but nothing assertive in datasheet on satisfying 704 requirements) and everything else via Glenair proprietary (with roughly half a quarter physical clearance between them) in what appears to be the equivalent of 38999 service class F finish (which won't survive 500-hr dynamic salt spray). I also do wonder if mechanical engineers would raise a nit at the countersunk retaining of enclosure top.

The hardest thing to swallow here is having to pay an undisclosed MIL premium for what will fundamentally be constrained by its consumer RPi CM4 core. Even if Curtiss-Wright lifetime buys for a gratuitous 10-yr support posture, it's hard to imagine how liabilities like tin whiskers can be pragmatically avoided, or what that means in terms of MTBF post-fielding.

This will probably fill someone's niche, but I agree that what's really missing is a robust RPi4 clone.

[1] https://www.curtisswrightds.com/products/computing/systems/s...

You can mitigate tin whiskers with a conformal coat application. It's not going to be as good as lead solder but it is an acceptable method when you have no other option.

Another possibility is that these are modified CM4 boards where parts were removed and resoldered using lead solder. The Broadcom chip is really the most critical component as you cannot buy them off the shelf but everything else could be purchased new and installed onto a stripped board or a brand new one made using the CM4 design files using lead solder.

> You can mitigate tin whiskers with a conformal coat application.

Precisely how effective do you think that'll be when your core packages are fine grid BGA?

> Another possibility is that these are modified CM4 boards where parts were removed and resoldered using lead solder.

This is the only long-term and standard solution I see, including package reballing. However, the fundamental problem with this approach is that the underlying PCB substrate is consumer grade that isn't designed for rework.

> wrapped in metal

Not even new metal: for good or bad, this “DuraCOR Pi” appears to be repurposing the chassis of the DuraNET 20-11 Ethernet switch: https://www.curtisswrightds.com/products/networking/switch-r....

> Before that, I'd honestly like to see TPM 2.0 support baked in first.

Maybe they could keep the price down by having a TPM connector and add the pinout for it, or even reuse some of the GPIO pins for it?

> the Pi CM4 may as well be vaporware - they're basically impossible to find anywhere even in single-digit quantities

That's near as damnit true for _all_ Raspberry Pi products right now. You can pretty much only buy any of them in single unit quantities, and it's like buying concert tickets, you need to keep your eyes open to find out exactly when someone has some for sale, and you need to jump on them before they sell out.

https://rpilocator.com

I managed to snag a Pi Zero 2 W a week or so back, had to get it shipped from Spain to Sydney which cost more than the Pi. I was limited to a single one at checkout.

Nice, it is currently available at the tiendatec.es shop, I managed to buy one. Shipping inside EU does indeed cost more than the board.

https://rpilocator.com/ is handy for anyone looking for them in stock anywhere.

It looks like they only search ~30 websites.

I just discovered my local RPi distributor has the 4B-2G in stock. (one less now, even though they're charging what looks like a fair but more than the usual "in stock in Australia" surcharge...)

https://raspberry.piaustralia.com.au

For me it is a joke, because the RPi has a long history of being an unreliable plateform.

Exemple: powercut and memory cards corruptions, etc...

So, if I had to design for military specs, I would probably bet on some other more robust platform.

MicroSD cards are IMO a poor choice to store the OS, using a USB thumbdrive or a USB enclosure is IMO more robust but not as compact.

Thankfully you can boot from those on the Pi.

I couldn't find any reference to storage options on the product page. The compute module is available with 0-32GB onboard eMMC but it's possible they're using an alternate storage solution on the carrier board. Hopefully they're using something more reliable than sd.

SD cards do seem to die very easily, from what I hear USB sticks are much more reliable.

From what I understood, it was not necessarily the quality of the SD card the problem, but that the rpi was lacking a capacitor or something like that preventing the current to be shut instantly when there is a power cut or a shutdown.

And sdcard don't like at all sudden loss of power in the middle of a 'write' or 'refresh' command.

Cards death or corruption happened to me a lot even when taking care with rpi but never when using beaglebones.

Would probably also need good performance on 128bit floats as well as ECC memory

I'm curious where 128bit floats become a significant concern? My understanding is that any critical numerical things would be done using integer arithmetic to eliminate any imprecision errors, and anything else likely is fine using doubles.

Integers are not often expressive enough for a lot of applications, so we use floats. I'm sure for a lot of things, doubles are good enough, but a better safe than sorry approach will lead you to use 128bit floats.

The case study that I learned in my Numerical Analysis course was the patriot missile incident, which lead to 28 death in the early 90s. Essentially, there was a timer implemented that used floats, and due to the inherent rounding error of floats, after enough iterations this caused the missile to miss its target by miles. There were some operational errors involved, such as leaving the missile "turned on" for longer than was suggested by the manufacturer, but still, these types of incidents should not happen.

Even with double precision, there are errors. For instance, 0.3 is actually represented as:

0.299999999999999988897769753748434595763683319091796875

Reply on news.ycombinator.com