gsuberland

gsuberland, 6 hours ago

@designthinkingcomic this is one of the things that bothers me about tetrapaks. they're greenwashed as being "made of recyclable materials" but it's a weird type of plastic-coated card with other plastic pieces attached, which can't be put in most recycling bins and which is actually pretty hard for recyclers to process. the nearest place to me that takes them is a 25 minute drive, and it's in the middle of an industrial estate in nowheresville so it's a huge pain in the ass.

gsuberland, 6 hours ago

@designthinkingcomic if I could hand a bag or crate of them to my supermarket delivery driver, that'd be convenient and easy. but that costs the supermarket money so they won't do it, despite selling hundreds of these things a day and 99% of them ending up in landfill.

gsuberland, 6 hours ago

what if someone pirates our unsolicited advertising video!? that'd be awful!

gsuberland, 15 hours ago

the TDR plot is the main interesting thing here. it plots characteristic impedance over time as an edge (wavefront) propagates through a PCIe 5.0 link.

nominal impedance should be 85Ω, but there are always variances in the traces and cables, and discontinuities at the connectors. the impedance mismatches cause insertion loss, and the link must meet a certain loss budget to function correctly.

immersion cooling adds a liquid dielectric which alters the impedance profile along the link.

gsuberland, 15 hours ago

@azonenberg der8auer / Caseking had the same issue with an aluminium contact frame recently. in the initial prototype the bottom of the frame was pressed against the PCB and caused a significant discontinuity in the top-layer traces, which caused enough loss to prevent most DDR5 modules from training. they milled just a few mm off the bottom and it worked again.

gsuberland, 18 hours ago

right now most of the power budget comes from the dielectric loss when getting the signal from the switching ASIC to the optical modules. most vendors have had to switch to multiple ASIC packages on boards to keep the link lengths short, and that's getting extremely expensive for 400G and 800G link speeds. the energy usage in driving those links is a big problem, too.

gsuberland, 18 hours ago

CPO constitutes a pretty decent win in terms of energy consumption (~40% reduction) but more than anything it's just useful from a cost and design standpoint because you're not having to deal with the complexities of moving TbE signals between the switch ASIC and the optics over copper. we're already well into the realms where controlled foil roughness on both the trace copper and back-side of adjacent ground planes is necessary, which ramps production costs a lot due to adhesion challenges.

gsuberland, 18 hours ago

but fully integrated photonics is a whole different ballgame. a 5x reduction in pJ/bit versus CPO, so that's an >85% reduction in energy consumption over current pluggable optics designs. wild.

gsuberland, 17 hours ago

I had been holding out hope that we'd see CPO adoption in general compute within the next decade or two, since we're already hitting major complexities in computing buses over copper (e.g. needing AFEs in the SERDES now) but honestly I reckon we might just skip right by it and end up with integrated photonics.

gsuberland, 16 hours ago

another fun quote from one of Intel's engineers, referring to the work they're doing with optical IO: "starting from where we were in early 2022, with 4Tbps over eight fiber pairs, we can see a clear path to doubling bandwidth at least 6 times without inventing new physics or dramatically changing the architecture, through increasing the line rate, wavelength count, and a one-time doubling through the use of polarisation".

that's 256Tbps on one package!

gsuberland, 16 hours ago

even more impressive is that the context of the quote is the bandwidth scaling they foresee based on the tech they've already managed to physically demonstrate in the lab. so the ring modulators, germanium photodiodes, lasers, optical amplifiers, fiber couplers, etc. can already perform at the necessary levels to achieve this stuff, it's just not ready for mass production.

gsuberland, 20 hours ago

from the same email

gsuberland, 20 hours ago

@thor yes. I joined a few and immediately regretted it. anything vaguely gaming or youtube related is a constant mess of arguments and pepe stickers. even machining / electronics stuff is just a mess of immature weirdos.

now I'm only in a few for events like Revision, NOVA, and my local hackspace. I'd prefer those to be on Slack but apparently Slack got really pushy with making you pay for everything.

gsuberland, 19 hours ago

@Natanael_L @thor yup, and also that it's really awful as a hosting platform for binary files, which is the primary thing forums like WSGF are using it for.

gsuberland, 19 hours ago

@Natanael_L @thor plus the core problem that if I have a transitive need for a single file or piece of information to solve an immediate problem, possibly the worst way to service that need is to subscribe to a persistent realtime communications feed about the general topic.

gsuberland, 12 hours ago

@JeffGrigg @fasterthanlime I'm not quite sure what you're trying to convey here. I have ADHD and most of the questions were applicable to ADHD.