Hacker News Reader

Comment by olabyne | Hacker News Reader

Does this mean the radiation hardening is a bit overkill for some applications ? I mean, even if the 801 Soc is really weak against radiations (apparently, not that weak for a 90day+ mission), it might be better to just throw 2 or 3 of them with redundancy instead of going for the good old RAD750 and its pound of weight, crazy cost and weak performances ?

Or is just crazy luck and the thing avoided the beams ? Might be plausible, if the last flight is not explained, it was the one flight with less radiation luck ?

View on HN

I know this is what most of the SpaceX mission control systems do. They're set up with ~3x/5x redundancy for effectively consumer hardware. If they detect memory misalignment between 1/3 of the nodes, they'll power cycle it while the other 2 take primary control. Big cost savings by not having to purchase radiation-hardened hardware and AFAIK they see similar reliability.

IIRC they have 3 pairs (6 total) and if any of the pairs have a mismatch they powercycle them both and let the other 4 continue as-is as you described.

Iirc, that's how the space shuttle control systems worked as well, but they were not COTS...

I believe the shuttles used five computers - four ran the same software and voted if there was a discrepency, whereas the fifth ran simpler independently developed software in case a common bug rendered all four primary computers inoperable.

Yes, most of the manned missions (at least Apollo and beyond as far as I'm aware) use the belt and suspenders approach of triple voting rad-hardened processors. So 3x RAD750s in a voting redundancy for main control systems.

I'm very curious if any analysis on this can be done and published, though I imagine it's difficult without access to the actual equipment.

Goddard does do a lot of radiation testing on commercial chips. I believe they usually publish the results and should be able to look it up.

Edit: database here, but I haven't looked for data on the chip: https://radhome.gsfc.nasa.gov/radhome/raddatabase/raddatabas...

Interesting data - I couldn't find the Snapdragon 801 ... but they have tested a nVidia Geforce 1050 (https://nepp.nasa.gov/files/29573/NEPP-TR-2018-Wyrwas-TR-17-...)

I wonder why they are testing GPUs? I would imagine there are at least some people somewhere writing papers about sending GPUs on rovers etc for better on-device AI/ML (e.g. image processing)

Also: brief mention of raspberry pi (no dedicated wreite-up like the GPU): https://nepp.nasa.gov/files/27888/NEPP-CP-2015-Campola-Paper...

Yeah I didn't find it there either, but there is an IEEE paper with testing [0]. Looks like they only did single event effect testing (SEE) and not total ionizing dose (TID) as well. The SEE is for looking and effects for individual memory upsets, latchup, gate rupture, etc. You can still do a lot with a chip if you are only getting the non-destrucrive effects with memory scrubbing and resets (as long as you're not mission critical). But the destructive effects (like gate rupture) are the ones that could make this totally infeasible depending on the limit.

Not surprised that they didn't do TID as it was only supposed to be a 90 day demo mission so not terribly long. And TID can be a costly test because it can take so long.

As to the GPUs, yeah the main application is some processing when you are comm bandwidth constrained. Then you can send down processed products or snapshots from something that is collecting a lot of data (like large imager or high bandwidth SDR)

[0] https://ieeexplore.ieee.org/document/8906649

Primarily for machine vision applications, both for robotic satellite servicing and for e.g. processing earth imagery data onboard to reduce downlink requirements.

Goddard does quite a lot of work on camera-based RPO (rendezvous and proximity operations) which require cameras and lidars to image a client spacecraft and calculate relative pose. This is the single most computationally taxing operation for robotic satellite servicing missions.

It stopped being about reliability twenty years ago. Now it's about lining pockets. The current top of the line RAD5545 is based on a chip from 2010 with 45nm lithography. It consumes 20 watts to perform 3.7GFLOPS. A Snapdragon 8 Gen 3 has a TDP of 12w and does 4.7TFLOPS. Over a thousand times the performance at less than a thousandth the price.

Radiation tolerance goes down with node size. Companies aren't pushing much smaller because they there's nothing they can do about the increased upset rates at those sizes.

Rather, they're turning to other options to achieve rad tolerance in critical systems, stuff like soft cores specifically designed for redundancy running on radiation hardened FPGAs.

Also, the rad tolerant parts aren't there to do the crazy high speed work. When people talk about voting architectures with redundant COTS parts, what do you think is counting the votes and resetting the effected systems?

Reply on news.ycombinator.com