And if you are curious about the modern radiation hardened CPUs then the current state of the art ones are the MOOG BRE440 [0] and the BAE RAD5500 [1], 5545 [2] being the highest performance multi core one.
Even more interesting that they both use the IBM POWER architecture!
There have been a number of rad-hard SPARC chips from different vendors tthat have flown along the way, and I know Frontgrade/Gaisler currently sells a SPARC v8 version, which isn't that far behind the 'state of the art' as the e5500 based PPCs from BEA (at least as far as state of the art in space rated, rad-hard processors goes...it's a conservative market). Quite a few rad-hard ARMs out there farther down the performance curve.
Frontgrade also advertises a rad-hard RISC-V, as does Microchip (a PIC64 variant), that I know nothing about, but seems like an inevitable next step. Seems like you could grab some Xilinx rad-hard FPGA and bobs your uncle.
> Back in the late 1970’s and early 1980’s Sandia National Laboratory (in Albuquerque NM USA) began building the capacity to design, fab, and test IC’s at scale (packaging was handled by Fairchild and Allied Signal).
We need more of this kind of thing, generally: government agencies building up in-house technical capability, instead of outsourcing everything to contractors.
For instance: there should be a government-controlled pharmaceutical manufacturer of last resort. The clear benefits would be to provide extra capacity and prevent things like Martin Shkreli's scams with Retrophin/Turing Pharmaceuticals (https://en.wikipedia.org/wiki/Martin_Shkreli#Thiola_price_hi...).
Sandia is not in house government, it is an FFRDC or federally funded research and development center, which are weird quasi governmental entities that are meant to avoid a lot of the pitfalls of in house government red tape, and the downsides of short term contracts. Think faster hiring/firing and higher pay than government, but longer term funding and goals than contractors. All the national labs are FFRDCs and there are a few others.
My dad worked for one for decades - the MITRE corporation, which is a nonprofit, and interestingly enough got the first .org domain. They also run CVE, and do a variety of R&D and legacy systems work, most notably air traffic control. They also entered vehicles in the DARPA challenges in the aughts, though did not do well.
Very interesting! Definitely some jargon I’ve not come across before.
“The chips were made on a n-on-n+ epitaxial substrate to provide latchup control, extensive guard rings around transistors were used and hardened oxides”
This is standard semiconductor manufacturing jargon.
"Substrate" here refers to the silicon wafer on which the integrated circuits are made, which at the end of the manufacturing process is cut into individual chips, which are then packaged as CPUs in this case.
An epitaxial wafer is a wafer on which epitaxial growth has been done before the rest of the manufacturing process. The wafers are cut from a huge crystal that has been grown from molten silicon. Initially they have a uniform concentration of doping impurities throughout their volume.
Epitaxial growth means that an extra layer of silicon is grown on the wafer and the growth is done in such a manner that all the layer is a single crystal and its lattice continues the crystal of the wafer, without interface defects.
The purpose is to have a different concentration of impurities in the extra layer, compared with the base wafer. N-on-n+ means that the initial wafer contained N-doping impurities, e.g. antimony, in a very high concentration (+), so that its electrical resistance would be minimum, while the "n" epitaxial layer also contains an N-doping impurity, e.g. phosphorus, but in a much lower concentration, so that it has a high electrical resistivity.
Both the fabrication of silicon wafers and the epitaxial growth are typically done by other companies than those that make integrated circuits, so the IC maker, or a silicon foundry like TSMC, buys epitaxial wafers according to a certain specification and they use them as the starting material in their manufacturing process.
"Latchup control" is a term specific to CMOS integrated circuits. In CMOS there exists a parasitic thyristor (a.k.a. SCR) composed of 2 parasitic bipolar transistors. If the parasitic thyristor turns on, it applies a short-circuit on the power supply, causing a huge electrical current spike, which normally destroys the integrated circuit, perhaps also other things if the power supply is not protected against short circuits.
In order to prevent the latchup of the parasitic thyristor, the structure is modified in various ways to reduce the gain of the parasitic transistors. If the gain is low enough, the thyristor cannot turn on.
Using a simple n substrate (which is cheaper) results in a high gain for the parasitic pnp bipolar transistor. Using an epitaxial n-on-n+ wafer reduces the gain of the pnp, lowering the probability of latchup.
Guard rings around transistors (which are made by diffusing certain doping impurities and then possibly also covering the diffused ring with a polysilicon or metal ring) have various purposes, typically related to preventing the electrical breakdown of the transistors at lower voltages than intended. This is especially important for radiation-hardened devices, because the most frequent effect of the passage of a ionizing particle through the semiconductor would be to generate mobile charge carriers that could cause the electrical breakdown of a transistor.
"Hardened oxide" is a more ambiguous term, but I assume that here it refers to high-quality oxide, i.e. which has a high value for the electrical field that can be sustained without electrical breakdown.
Interesting combination of 'remarkable' and 'wtf' that we fling nuclear weapons around with the computational equivalent of a couple of TRS-80s[1]. I can only imagine the sighs of relief from the devs when things like the MIL-STD-1750a and later rad-hard SPARC and PPC variants came along.
[1] yes...I know the TRS-80 had a z80, not an 8085. Close enough.
On the flip side, the fact that those processors were enough to steer spacecraft make me feel like there’s also a decent amount of remarkable wtf in how much compute we have now and how little we get out of each instruction on average compared to what people were doing with these z80 equivalents.
When you don't have the overhead of an operating system with decades of backward compatibility cruft, a scheduler, a virtual memory controller, and a file system you can accomplish amazing things with simple processors. Bare metal is something I'd encourage every programmer to try.
I actually have a 8085 Primer Trainer that I built/used in the 90s for school. [1]
Programmed with pushbuttons and 7 segment LEDs. No backup memory so if you shut it off it lost it's program and you had to start over.
Not a programmer by trade, I prefer hardware...had no idea until recently how valuable the training was. We learned BASIC and 8085 machine code as well as building logic circuits from discrete parts. Then I used basically no code myself for 15 years until I learned Arduino.
Knowing the basics certainly helped me know what was going on. From there it was just syntax for languages.
Yeah exactly, we now have so many layers of stuff. On top of vmem & OS, add high def displays, and today’s corporate firewall and malware scanning. I wouldn’t be surprised if just booting my Win 11 laptop, logging in, and launching Teams uses more compute than the entire Galileo mission used over its entire 8 year run. :P
Even without the layers & cruft though, the raw perf is astounding to those of us who remember 8 bit 1Mhz microprocessors. Today’s gamers are used to double-digit teraflops(!) of compute, just to render all the pixels for Minecraft or Fortnite.
I don’t know if there’s a better way these days, but for me Arduino has been an easy & super fun way to futz with a tiny bare metal microprocessor.
Even hypersonic weapons with precision terminal guidance use truly ancient CPUs. Physics limits of molecular materials places a very low upper bound on the amount of compute required.
The rate at which an object in the physical world can alter its trajectory is ultimately limited by the strength of molecular bonds in the material it is made from. Exceed that limit and the object will disintegrate. This upper bound is extremely slow from the perspective of a CPU, making it computationally trivial. A computer can react orders of magnitude faster than the quickest physical objects.
I imagine you're correct about course correction speed, although I'd also expect that the materials and their properties are quite classified at present.
I would also imagine that there could be processing necessary that is mostly unrelated to manoeuvering speed (inlet/control surface management, etc). Perhaps some hypersonic experts could weigh in and let us know :)
The theoretical limits of materials science is undergrad chemistry. Among other things, the bond energy between atoms in a molecule tells you how much energy is required to break them. The strongest possible bond is less than 3x the energy of the ubiquitous carbon-carbon bond of organic chemistry. There is a bit more to it than this e.g. the lattice structure also matters, but you get the gist. You can specify an energy that is not particularly high that is known to exceed that required to destroy all molecular materials.
Real materials are considerably worse than theoretical and you also have to optimize for other properties. The classified part is often more about the processes for fabricating specific materials. This can be extremely non-obvious and difficult.
All the exposed surfaces of a hypersonic missile are themselves limited by material physics. A major problem in hypersonic research was material surfaces degrading so quickly that they lose precision of control. The CPU can send commands but the physical materials of the guidance system don't respond in a reliable way due to the material degradation.
If this was copy-pasted, isn't it much more likely that it was copy-pasted from a document describing the performance of the SA3000 chip, than from a document that was written before the SA3000 was developed?
The only overlap from the document with the text you quote is the "106", which is a pretty common mis-formatting issue.
Bad typesetting just indicates poor editing, not slop IMHO.
I guessed after about a second of thought that this actually meant 10^6. If I had to guess how this happened, somebody just wrote their prose in Word with the 6 in superscript and cut and paste it into HTML which lost the formatting.
I think it's a shame if you consider the whole article (which I personally found very interesting) as slop because of 2 incorrectly formatted numbers that could easily result from a cut-and-paste error. It's clear the article hadn't been well proof read as there are a number of spelling mistakes too, but that doesn't make the content itself slop.
Like others have mentioned, 50k small CPUs like 8085 can be made in a single production batch (i.e. a small basket containing silicon wafers, which passes through all manufacturing steps), so a number like this is likely to be the minimum amount that can be produced.
The customer would order this minimum quantity, and most of it will probably be kept as spares.
That number was probably shaped by minimum production-run requirements, alongside the need for software development units, along with other factors, like the use in Trident II and other quests we may not know about.
There isn't really a minimum production run for silicon chips, they do small test runs all the time to test new designs.
At least not from a practical perspective.
From an economic perspective, stopping after a single small run is just wasteful. The upfront design costs are so high, and the per wafer costs are so slow that you might as well make a lot extra. Maybe you can find a use for them, or sell them to someone else.
Trident 2 (article says used 8 of these chips), and google says around 400-424 made, so easily would have soaked up 4000 of these CPUs with spares alone. So if anything, the production run seems light.
I read they had their own fab, so the minimum production run aspect would appear moot.
I assume the tooling and process are such that it’s a one-time thing, as in, this is the most of these chips that we could ever possibly need for all time. They’re not going to be able to spin up the same fab and build the same chips the same way again in the future whether that’s 5 or 50 years in the future. Given the long lifespans of military systems, it’s maybe not so crazy.
Even more interesting that they both use the IBM POWER architecture!
0, https://www.moog.com/products/avionics/spacecraft-avionics/b...
1, https://en.wikipedia.org/wiki/RAD5500
2, https://web.archive.org/web/20190226111129/https://www.baesy...
Frontgrade also advertises a rad-hard RISC-V, as does Microchip (a PIC64 variant), that I know nothing about, but seems like an inevitable next step. Seems like you could grab some Xilinx rad-hard FPGA and bobs your uncle.
In comparison radtherapy patients get 20 gray in 1-2 weeks so it's the 20/10000 = 0.02% of what these designs target
A whole body exposure of 5Gy would kill you, although it would happen over two weeks.
10kGy? Spectacularly fatal.
We need more of this kind of thing, generally: government agencies building up in-house technical capability, instead of outsourcing everything to contractors.
For instance: there should be a government-controlled pharmaceutical manufacturer of last resort. The clear benefits would be to provide extra capacity and prevent things like Martin Shkreli's scams with Retrophin/Turing Pharmaceuticals (https://en.wikipedia.org/wiki/Martin_Shkreli#Thiola_price_hi...).
My dad worked for one for decades - the MITRE corporation, which is a nonprofit, and interestingly enough got the first .org domain. They also run CVE, and do a variety of R&D and legacy systems work, most notably air traffic control. They also entered vehicles in the DARPA challenges in the aughts, though did not do well.
But then how will politicians favor their campaign donors?
“The chips were made on a n-on-n+ epitaxial substrate to provide latchup control, extensive guard rings around transistors were used and hardened oxides”
"Substrate" here refers to the silicon wafer on which the integrated circuits are made, which at the end of the manufacturing process is cut into individual chips, which are then packaged as CPUs in this case.
An epitaxial wafer is a wafer on which epitaxial growth has been done before the rest of the manufacturing process. The wafers are cut from a huge crystal that has been grown from molten silicon. Initially they have a uniform concentration of doping impurities throughout their volume.
Epitaxial growth means that an extra layer of silicon is grown on the wafer and the growth is done in such a manner that all the layer is a single crystal and its lattice continues the crystal of the wafer, without interface defects.
The purpose is to have a different concentration of impurities in the extra layer, compared with the base wafer. N-on-n+ means that the initial wafer contained N-doping impurities, e.g. antimony, in a very high concentration (+), so that its electrical resistance would be minimum, while the "n" epitaxial layer also contains an N-doping impurity, e.g. phosphorus, but in a much lower concentration, so that it has a high electrical resistivity.
Both the fabrication of silicon wafers and the epitaxial growth are typically done by other companies than those that make integrated circuits, so the IC maker, or a silicon foundry like TSMC, buys epitaxial wafers according to a certain specification and they use them as the starting material in their manufacturing process.
"Latchup control" is a term specific to CMOS integrated circuits. In CMOS there exists a parasitic thyristor (a.k.a. SCR) composed of 2 parasitic bipolar transistors. If the parasitic thyristor turns on, it applies a short-circuit on the power supply, causing a huge electrical current spike, which normally destroys the integrated circuit, perhaps also other things if the power supply is not protected against short circuits.
In order to prevent the latchup of the parasitic thyristor, the structure is modified in various ways to reduce the gain of the parasitic transistors. If the gain is low enough, the thyristor cannot turn on.
Using a simple n substrate (which is cheaper) results in a high gain for the parasitic pnp bipolar transistor. Using an epitaxial n-on-n+ wafer reduces the gain of the pnp, lowering the probability of latchup.
Guard rings around transistors (which are made by diffusing certain doping impurities and then possibly also covering the diffused ring with a polysilicon or metal ring) have various purposes, typically related to preventing the electrical breakdown of the transistors at lower voltages than intended. This is especially important for radiation-hardened devices, because the most frequent effect of the passage of a ionizing particle through the semiconductor would be to generate mobile charge carriers that could cause the electrical breakdown of a transistor.
"Hardened oxide" is a more ambiguous term, but I assume that here it refers to high-quality oxide, i.e. which has a high value for the electrical field that can be sustained without electrical breakdown.
[1] yes...I know the TRS-80 had a z80, not an 8085. Close enough.
https://en.wikipedia.org/wiki/Apollo_Guidance_Computer
Not a programmer by trade, I prefer hardware...had no idea until recently how valuable the training was. We learned BASIC and 8085 machine code as well as building logic circuits from discrete parts. Then I used basically no code myself for 15 years until I learned Arduino. Knowing the basics certainly helped me know what was going on. From there it was just syntax for languages.
[1] https://flic.kr/p/2mkG7gC
Even without the layers & cruft though, the raw perf is astounding to those of us who remember 8 bit 1Mhz microprocessors. Today’s gamers are used to double-digit teraflops(!) of compute, just to render all the pixels for Minecraft or Fortnite.
I don’t know if there’s a better way these days, but for me Arduino has been an easy & super fun way to futz with a tiny bare metal microprocessor.
The rate at which an object in the physical world can alter its trajectory is ultimately limited by the strength of molecular bonds in the material it is made from. Exceed that limit and the object will disintegrate. This upper bound is extremely slow from the perspective of a CPU, making it computationally trivial. A computer can react orders of magnitude faster than the quickest physical objects.
I would also imagine that there could be processing necessary that is mostly unrelated to manoeuvering speed (inlet/control surface management, etc). Perhaps some hypersonic experts could weigh in and let us know :)
Real materials are considerably worse than theoretical and you also have to optimize for other properties. The classified part is often more about the processes for fabricating specific materials. This can be extremely non-obvious and difficult.
All the exposed surfaces of a hypersonic missile are themselves limited by material physics. A major problem in hypersonic research was material surfaces degrading so quickly that they lose precision of control. The CPU can send commands but the physical materials of the guidance system don't respond in a reliable way due to the material degradation.
Fun problems.
What, so that they can debug in Chrome and put the fusing and inertial navigation processes in isolated web views?
The inertial navigation system is the very crazy part, along with the nuclear fusion warhead design itself.
https://youtu.be/AazmxNs5kmE?is=2LE2q3rBSWDyTs7j
https://thebulletin.org/2017/03/how-us-nuclear-force-moderni...
Clearly you meant the TRS-80 Model 100.
> An 8085 processor that could handle 1×106 rads of radiation with only a 25% reduction in performance, and 3×106 rads with a 40% drop.
Hmm, from where did they copy-paste this mangled scientific notation?
Ah here we are, pg. 37 (46 in PDF file): https://apps.dtic.mil/sti/tr/pdf/ADA063902.pdf
The only overlap from the document with the text you quote is the "106", which is a pretty common mis-formatting issue.
I guessed after about a second of thought that this actually meant 10^6. If I had to guess how this happened, somebody just wrote their prose in Word with the 6 in superscript and cut and paste it into HTML which lost the formatting.
I think it's a shame if you consider the whole article (which I personally found very interesting) as slop because of 2 incorrectly formatted numbers that could easily result from a cut-and-paste error. It's clear the article hadn't been well proof read as there are a number of spelling mistakes too, but that doesn't make the content itself slop.
I seriously doubt you need to fabricate 50k CPUs for a single space probe, including backups, testing chips, etc.
The customer would order this minimum quantity, and most of it will probably be kept as spares.
At least not from a practical perspective.
From an economic perspective, stopping after a single small run is just wasteful. The upfront design costs are so high, and the per wafer costs are so slow that you might as well make a lot extra. Maybe you can find a use for them, or sell them to someone else.
I read they had their own fab, so the minimum production run aspect would appear moot.
Back then an interface between terrestrial computer systems and a Zeta Reticulan spacecraft required a small supercomputer on our side.