Like, very interesting article. It also mentions bit flips.
I think the major protection against bit flips is to use larger structure sizes. Like, today’s desktop computers use structure sizes around 2 nm or sometimes even less than that. Which means you only need to shift a few electrons around to cause a bit flip. If you use larger wires inside the processor, they store more electrons so it takes more energy to flip them around. So there’s a much much smaller number of radiation particles that have a minimum of that energy, so lower risk of bit flips.
Does anybody have actual numbers on the structure sizes needed to effectively prevent bit flips? I mean, outer space missions like NASA flights already need this today.
Some newer radiation hardened stuff is 10x larger than that, older gear even more so. But that just reduces the risk, not sure it’s possible to negate it entirely.
An easier way is to just include more CPUs as part of the system, run them in lockstep, then compare the results by majority rule. If 2/3 CPUs say one answer and the third says something else, you discard the result of the third and go with the other CPUs.
Like, very interesting article. It also mentions bit flips.
I think the major protection against bit flips is to use larger structure sizes. Like, today’s desktop computers use structure sizes around 2 nm or sometimes even less than that. Which means you only need to shift a few electrons around to cause a bit flip. If you use larger wires inside the processor, they store more electrons so it takes more energy to flip them around. So there’s a much much smaller number of radiation particles that have a minimum of that energy, so lower risk of bit flips.
Does anybody have actual numbers on the structure sizes needed to effectively prevent bit flips? I mean, outer space missions like NASA flights already need this today.
Some newer radiation hardened stuff is 10x larger than that, older gear even more so. But that just reduces the risk, not sure it’s possible to negate it entirely.
An easier way is to just include more CPUs as part of the system, run them in lockstep, then compare the results by majority rule. If 2/3 CPUs say one answer and the third says something else, you discard the result of the third and go with the other CPUs.