dr.-ing. ⁡jaseg

last night, a friend and I recycled some old PCBs left over from another project into notebook covers! We cut the pages and stitched the spine by hand. #electronics #bookbinding #diy

edit: it has found a new home! Does anyone here need or want an old 12”ish Philips TP200 amber CRT monitor that has a broken power supply? It’s yours if you pick it up or cover shipping from Berlin. #retrocomputing #electronics

@gsuberland@chaos.social the architecture is pretty nuts, you have 125 GS/s over something like JESD204 into a dedicated chonky HBM FPGA per channel pair, but that’s then only hooked up to the outside world through a PCIe 3 x4 interface at ~~30 Gbit/s. I suspect the vintage epyc3000 CPU is the limiting factor here, I’d expect the FPGA to do at least PCIe 4 here.

I have zero need for a mechanical typewriter but I found this one for free on the street so I couldn’t really walk past 😅 now I can finally shitpost in style

@gsuberland@chaos.social also that price, at several k$ just for the acquisition FPGA and four of them being needed for eight channels you’re quickly into the 100 k$ MSRP range once you apply a typical profit margin for a new specialty product like that.

@gsuberland@chaos.social with the way this tech is going, I’m kind of wondering when they’ll drop the pricy FPGA and just hook up the ADC directly to the PCIe fabric on a recent Epyc and just stream straight to main memory. On a recent epyc with all memory channels populated a quick back-of-the-envelope calculation comes out to enough capacity to run 4x125 GSp/s on a single CPU.

@uriel@bbs.keinpfusch.net you sound like every other guy who thinks the natural sciences are somehow superior to humanities, which is an anti-intellectual position. As someone with a degree in some of those sciences you seem to like, I can say that the work of Bender, Gebru et al. is entirely valid, and surpasses most of what I see published in other parts of this science in its merits. @teledyn@mstdn.ca

@whitequark@social.treehouse.systems @karotte@chaos.social @aryl@mastodon.social every household should have a thermal camera to check on drying laundry and a laser welder to properly reseal partially used cans of beans and such.

Has anyone ever noticed how it's really emailGPT rather than chatGPT with how verbose these things are?

@Lunaphied@provably.online sorry, I forgot to link. Here’s the presentation that is from: https://indico.cern.ch/event/1606914/contributions/6771542/attachments/3178052/5652347/2025.11.19_SAMBA_HPTD.pdf and this is the working group this came from. They publish a lot of stuff here when you click through the links: https://indico.cern.ch/category/10229/

They wrote up a design doc at the URL below. The gist of it is that they chose a fixed size and error correction level, then unrolled the whole encoding process into a finite circuit, and implemented RS encoding as a linear transformation over GF(2). In effect this is similar to what you'd do in digital circuit design, or mapping an algorithm to a secure multiparty computation implementation. https://github.com/jimparis/qr-font/blob/master/design.md

@netspooky@haunted.computer I wrote a paper a while ago where we were considering a scenario of using a datacenter as a controllable electronic load for some experiments. We figured it would be pretty hard to actually break anything, even at much larger scale. I think in practice, if someone were to try what they describe in that paper, it would hardly have an impact beyond the immediate vicinity of the datacenter. https://dl.acm.org/doi/pdf/10.1145/3564625.3564640

@netspooky@haunted.computer It's a bit funny that in their paper they talk about EMI as a side channel, which would only work across short distances. What we showed in our paper (through simulation) is that if you have access to a big enough load like this, modulating it properly you can broadcast information across the whole continent. If your load is smaller, using local phase variations I'd guess you could still broadcast across a few hundred km radius without much trouble.

@doctormo@floss.social take your time and have a chill sunday! This sort of thing can wait until next week. The world can do a sunday without inkscape.org

@drmorrisj @hannorein @daisy in my experience in teaching in higher education, it makes a big difference how easy cheating is. LLMs made cheating easy enough that it went from maybe 20% of participants in a first semester lecture to something like 50%. Those additional 30% of students would have learned something before, but are near guaranteed not to learn anything now.

@karotte@chaos.social @whitequark@social.treehouse.systems @aryl@mastodon.social a similar cost reduction also happened with laser engravers and laser welders. I think those now are at the point where they start saving on stuff you might actually want.

@xssfox@cloudisland.nz @haakon@aus.social @netspooky@haunted.computer In the paper, I think I was using a chip rate (DSSS modulation) of 1 Hz (2 Bd). You can’t easily go much faster because there’s lots of nose below ~1 Hz decreasing SNR and at least in the ENTSOE you’d have to whiten your transmission to avoid feeding the global resonance modes. Beyond that, it’s just a tradeoff of transmit power vs. bandwidth. I assumed lower power, and made up for it in coding gain.

@Gyroplast@furry.engineer an OCXO will give you much better phase noise than a PLL, but the difficulty is synchronizing them. In these experiments, they have hundreds or thousands of boards synchronized to a common reference clock.

@cccpresser@chaos.social @janamarie@mystical.garden @feuerrot@chaos.social gerbolyze / gerbonara can do lossless svg to gerber. I’m happy to help if you have questions :)

@azonenberg@ioc.exchange I’m coming to that same realization now :)

@karotte@chaos.social and it’s confidential!

@whitequark@social.treehouse.systems combined with some sort of pipe/endpoint multiplexer you could use that to make glasgow into a monster multichannel UART adapter

@coral@empty.cafe I think the scenario of this experiment was that they already had a large number of boards deployed that were lacking “proper” hardware capabilities for this kind of fine-grained adjustment and they evaluated retrofitting that capability to these exising boards in software.

@xssfox@cloudisland.nz @haakon@aus.social @netspooky@haunted.computer that sounds fun!

@StripeyYena@chaosfurs.social pretty!

@tisba@ruby.social Sorry! I didn't mean to argue! Myself, include I think most people working in that area agree that in a better world, we would have written all media decoders in memory-safe languages. I just wanted to illustrate why this kind of thing can still happen today.

@tisba@ruby.social media decoders in particular are a bit thorny. They're extremely complex, so nobody bothers to rewrite them in a memory-safe language. They are also highly performance sensitive and often carry around piles of hand-optimized assembly, which is unsafe as is. Finally, they often interface with hardware that does part of the decoding, and all bets are off as far as the memory-safety of that is concerned.

@tisba@ruby.social @cwebber@social.coop tbh media decoding always was a huge attack surface. If you're really worried about this stuff, QubesOS provides a solid solution that would prevent any chance of RCE or LPE by this or any other decoder bug. For applications like thumbnailing potentially untrusted files, solid sandboxing should absolutely be employed. Since thumbnailing is a simple bytes -> bytes operation with no need to perform system I/O, you can build a robust sandbox around it.

@krnlg@mastodon.social Here’s the presentation where that is from. In that presentation they introduce a new clocking architecture for future systems that allows fine phase control in a more direct way by chaining two fractional-N PLLs: https://indico.cern.ch/event/1606914/contributions/6771542/attachments/3178052/5652347/2025.11.19_SAMBA_HPTD.pdf

@krnlg@mastodon.social the sub-heading says they control temperature using the fans in order to phase shift the clock. They measure and lock against an externally supplied, very stable reference clock. Their application is an existing board that has a clock tree that provides some phase/frequency control, but that control isn’t as fine grained.

@jhostert@mathstodon.xyz it’s in principle similar to an ovenized quartz oscillator, just that this system was never designed for this.

@davidr@hachyderm.io about half an hour I think. The original slides are linked down thread.

@John_Ligher@urusai.social I think if you can avoid it there are better ways to achieve the same result. AFAIU this was meant as a retrofit for boards that lacked the necessary circuitry to do this in a nicer way.

@coral@empty.cafe here you go: https://chaos.social/@jaseg/116567746676154596

@gosha@merveilles.town yeah, usually whois. Often it’s just abuse@whatever.com.

@gosha@merveilles.town I’m using fastmail too and I found their spam filter to be generally ok but sometimes oddly insensitive to certain particular types of spam. Usually, what gets through is spam with a valid origin server in a reputable location. In these cases, I usually escalate straight to their hoster’s abuse contact and train my a client-side filter on the mail.

@azonenberg@ioc.exchange I’m going to have to upgrade my setup some time in the next 1-2 years. I kind of hope that by the time it becomes really necessary, the bubble will have deflated by a bit.

@haakon@aus.social @netspooky@haunted.computer IIRC it was between 10 and 100 mBd (that’s milli, not mega) @xssfox@cloudisland.nz

@azonenberg@ioc.exchange yeah right now, that’s still pretty decent and apart from GPUs and 100GBE-class network adapters there isn’t much that should saturate that. For an ssd, gen3 is plenty as long as it’s the full four lanes.

@azonenberg@ioc.exchange Yeah I really want to avoid buying more hardware, and also my home office situation doesn’t really work with loud server hardware. When I upgrade, I think I’d go for something that I can fit into a desktop case (ATX/EATX with an older server or workstation platform)

@skyguided@cathode.church I have a cheapo one that amazingly just works for 5 years and counting. I’ve had to replace a broken cable once and I put in quieter fans, but overall I’ve done lots of prints with almost no messing about with the printer.

@littlealex@infosec.exchange awesome, are you local or do you need it shipped?

i need more pcie lanes

@azonenberg@ioc.exchange I have one of those doing x8 -> 2x x4 m-key. I'm mostly limited by my (consumer) mainboard only having ~24ish lanes total, so there's only one full x16 slot anyway.

@azonenberg@ioc.exchange at these capacities I wonder whether the 200-400 USD for one of those PCIE/m.2 cards that have an onboard PCIE switch might be worth it. Some of them only do 2.0, but I have seen some that advertise 4.0 and above.

@mttaggart@infosec.exchange I think the primary failure case of any such digital “proof of humanity” system will be something you could call an arbitrage gap: As soon as these digital proofs become useful enough to carry any value, bad actors will find ways to force or bribe poor people into providing them with their proofs. Given the enormous wealth inequality across the globe, I suspect this sort of identity arbitrage would scale very well.

@azonenberg@ioc.exchange ah, that makes sense.

@azonenberg@ioc.exchange this! I recently got this really janky x4 extender with a tight rect-angle plug to bring out the one remaininng x4 slot that's currently blocked by my dual-slot GPU to hook up a SSD to it.