Security and open source at the hardware/software interface. Embedded sec @ IOActive. Lead dev of ngscopeclient/libscopehal. GHz probe designer. Open source networking hardware. "So others may live" Toots searchable on tootfinder.
Public Key
npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Profile Code
nprofile1qqsvxk504cz6sw9cvdc8sqamrl0uljanntw0q8hp48799m6ynxxwfsqpz3mhxue69uhhyetvv9ujuerpd46hxtnfduqs6amnwvaz7tmwdaejumr0ds75w78g
Show more details
Published at
2026-08-17T09:28:56Z Event JSON
{
"id": "3ae6dc84756b1a2cd852dab6a4168319792eece73b1b43118926509963dbbab8" ,
"pubkey": "c35a8fae05a838b863707803bb1fdfcfcbb39adcf01ee1a9fc52ef44998ce4c0" ,
"created_at": 1786958936 ,
"kind": 0 ,
"tags": [
[
"proxy",
"https://ioc.exchange/users/azonenberg",
"activitypub"
],
[
"client",
"Mostr",
"31990:6be38f8c63df7dbf84db7ec4a6e6fbbd8d19dca3b980efad18585c46f04b26f9:mostr",
"wss://relay.ditto.pub"
]
],
"content": "{\"name\":\"Andrew Zonenberg\",\"about\":\"Security and open source at the hardware/software interface. Embedded sec @ IOActive. Lead dev of ngscopeclient/libscopehal. GHz probe designer. Open source networking hardware. \\\"So others may live\\\"\\n\\nToots searchable on tootfinder.\",\"picture\":\"https://files.ioc.exchange/iocexchange/accounts/avatars/109/320/697/263/826/535/original/452045644aa6671b.jpg\",\"banner\":\"https://files.ioc.exchange/iocexchange/accounts/headers/109/320/697/263/826/535/original/e9bac12e369b65ab.jpg\",\"nip05\":\"[email protected] \",\"fields\":[[\"ngscopeclient\",\"https://www.ngscopeclient.org/\"],[\"Blog\",\"https://serd.es\"],[\"Location\",\"Seattle area\"],[\"GitHub\",\"https://github.com/azonenberg\"]]}" ,
"sig": "cb95ac608aeab68392a1b7690fbb84c99f889d01b555a8af01eb4e9c9dc72da571cf1bb2c7079e4e3888eeaa0fe937637ca6fff6a4671d09d51eb914990f5ed7"
}
Last Notes npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg The design target for this filter was 10.25 GHz Fc and 250 MHz -3 dB BW. The actual as-fabricated filter seems to be more like 9.92 GHz Fc and 292 MHz -3 dB. I don't really feel confident making any claims about stopband rejection because of the poor dynamic range of this setup, but the hump around 10.7 GHz is not ideal. I saw some ground plane resonances in the simulation and wasn't sure if they were real, it seems they might be? Once I get more detailed S-parameter data with better dynamic range I'll go back to the simulations and revisit. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…t7ks so i just wanted it to function long enough to smoke test npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…cllx @nprofile…huwm 7.4 tflops fp64 is quite good, although volta is ancient enough NSight probably doesn't support it anymore. Also i retired a kepler quadro recently because the driver no longer compiled on a current Debian kernel. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…huwm I've talked to the Sonnet people about GPU accelerating their solver as well. It's definitely something I want to think about. One of the big challenges is consumer GPUs having poor float64 performance, and a lot of scientific computing stuff really needing the extra dynamic range from fp64. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg I'm offering my oscilloscope probing class in Europe for the first time! If you're in or near the Netherlands, consider signing up. Space is limited to 10 students to ensure everyone has enough lab time; two spaces are already taken. https://www.eventbrite.com/e/oscilloscope-probing-training-tickets-1996971622881?aff=oddtdtcreator npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…ptud is it bad that I know it's a 3m 6000 series instantly? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…skdr It is a huge pain in the butt to configure and I really don't like the architecture. They're also close to $100 in low volume and lack a bunch of features the STM32s have that I consider key for useful embedded stuff ,most notably a hard TRNG and AES accelerator npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Once again daydreaming about a "baby zynq" microcontroller with a STM32H7 class processor and memory subsystem attached to a few kluts of low power FPGA with hard APB/AHB/AXI (depending on how big they want to make it) ports, a bunch of dedicated or muxable GPIOs, etc. ST apparently has an FPGA-esque peripheral in some STM32s but it's *sixteen* LEs which is a joke. You can't do any kind of useful custom peripheral in that. Ideally i would want to be able to implement my own peripherals for a weird bus the MCU doesn't include natively and even have native DMA to/from them. Which means having full MMIO capabilities (no need to put a DMA in the FPGA, you can use the hard DMA IP and just read/write registers in the FPGA). npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…3tsk also i think you have to add -c:v libx264 before the -x509 npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…3tsk shitposts for a very specific audience npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…k7h5 @nprofile…skdr I cannot even load the analog.com home page npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg For decades there's been a conflict between governments wanting to get access to people's private data and people trying to keep them out. Over the years I've gradually come to the conclusion that the stable "Goldilocks zone", where there's the least likelihood of a major swing that makes things worse for everyone, is *not* the privacy utopia where devices are invulnerable, or a police state where anyone with a badge who asks can root through all your data, or a fictional engineered cryptographic backdoor that''s somehow only exploitable by "the good guys". Instead, it's the state where devices (especially phones) are breakable, but only through attack vectors that require physical access, have a high per-device cost, and are not 100% reliable. Things like "solder to an internal layer trace on the logic board to glitch it" or "FIB a per-device key out of the SoC". I'm not saying we should be adding deliberate backdoors - quite the opposite, since these are very likely to make entry *too* easy. But if devices become completely invulnerable, I expect we will see something like chat control turned up to 11 where the pressure to find some way in becomes unbearable and we end up with CALEA-style entry points that turn devices into total spyware with no privacy at all. So IMO the ideal balance is where there is no viable remote attack vector, trivial drive-by USB etc attacks a la Cellebrite are blocked, but if a government has a phone off a body and are willing to spend $100K+ of lab time to have a >50% chance of breaking into *that one device*, destroying it in the process, without it easily scaling to the next, they can do so. And this level of difficulty is juuust easy enough that they aren't spending the same money lobbying politicians to make it easy to break into every device belonging to an opposition party or something. They still try of course, they always want more, but they can't claim to congress that the guy who shot senator X will never be caught because they couldn't break the phone they found at the scene. You know your device was confiscated (or are dead and not around to complain), it's not viable to do to everyone crossing a border or who was in a certain geofenced area, etc. There's no practical way to scale physical attacks to mass surveillance. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…mdnh this is using electric fields to propel atmospheric gas particles as reaction mass? Like an ion rocket crossed with a ramjet? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Just set up some new IAQ monitoring stations around the house. Still want to put one in the bedroom to see how stuffy it gets at night with the door closed, but I only had four units so this is what I started with. Particle counts are Sensirion SEN5x, CO2 and most other atmospheric data SCD4x. Haven't done any kind of zero cal on the CO2 and they've been sitting around idle for a while, so there's probably a bit of offset that will need to be corrected - the one on the deck is reading a bit higher than I would expect from outdoor air https://files.ioc.exchange/iocexchange/media_attachments/files/116/950/564/041/956/857/original/9632329238037a9f.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…3q9y pixel 9 pro and i see clouds npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…94p0 HBM is mounted to the chip substrate not the PCB, it's not really possible to rework practically unless you're making your own silicon interposer and even if you did I'm not sure how you'd extract it without destroying it. If you can reuse the whole chip package including the HBM and whatever it's attached to different story. But you can't like put it on a DIMM or something. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Huh, my local library's little news magazine is no longer called "Inspire". I guess someone finally told them about the naming collision with the al-Qaeda publication? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…kuux As of right now, we have: 1) Farberware toaster oven at my place with no controller that I've profiled and works well manually 2) Toaster oven at their shop with a little MCU based controller 3) Actual Essemtec conveyor oven that isn't in service yet because it needs 3-phase power and the current shop doesn't ave that, but it was too cheap on eBay to pass up npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Hmmm. Would it be accurate to interpret Manchester coding as a degenerate CDMA with a chip rate of 2x the baud rate and a spreading sequence of 0101? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg ILA is now working, but only for single-bit signals. Multi-bit still needs a data type defined in ngscopeclient (yes, we've gone this long without defining a multi bit waveform type) https://files.ioc.exchange/iocexchange/media_attachments/files/116/863/586/924/300/868/original/7801c99e7abea9ad.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…6cfx I plan to be building out a bunch of DP stuff in ngscopeclient soon if that will be helpful for you? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Chat what kind of fiber connector is this https://files.ioc.exchange/iocexchange/media_attachments/files/116/840/757/427/374/824/original/3d67f9e9c120836c.jpg https://files.ioc.exchange/iocexchange/media_attachments/files/116/840/757/860/604/210/original/5010bff56b06a7b6.jpg https://files.ioc.exchange/iocexchange/media_attachments/files/116/840/758/209/634/727/original/b37dabcf355e6307.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…u57d but like this Ti375 has 1344 DSP tiles that can each do one 19x18 integer + 48 bit accumulate, or a BFLOAT16 FMA, per clock at 1 GHz in the -4 speedgrade. So that's 1.34 TFLOPs of float16 or 1.34 TMAC/s of integer performance. Nothing to sneeze at, but the AMD R9700 in my desktop workstation will do 95 TFLOPs of float16 (or 47 TFLOPs of float32 which is more useful to me - I'm not sure if you can gang the efinix DSPs to do float32 or not) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…u57d I tend not to do much number crunching on FPGAs, I use them mostly for high speed digital packet datapaths and data shuffling between network interfaces, ADCs, etc. Now, GPUs? Those are number crunching monsters lol npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…vkla It seems more like weird analog quirks where they have two different equalizer algorithms for different data patterns, and somehow there was a bug where they tied that to the datapath width somehow npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Cursed. I wonder how they figured *that* out https://files.ioc.exchange/iocexchange/media_attachments/files/116/838/754/632/362/938/original/f4ba117affca1656.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Setting up the Efinix devkit so I can bring up Ethernet on the SFP+ interfaces and sniff with my big scope to validate. Do you think I have too many optics? Lol https://files.ioc.exchange/iocexchange/media_attachments/files/116/838/156/553/299/224/original/cc165efc570648e5.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…e6sg the STM32 can only do gen2 x1 while the efinix devkit can do gen4... x4 I think, but that's plenty for my use case, I just want a SoC connected to an FPGA via a simple memory mapped interface with a minimum of IOs. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…e6sg yeah mini pcie supports usb instead of pcie over the same connector (as does M.2). I don't plan to use it. Use case is connecting the Efinix Titanium PCIe devkit to the STM32MP257F-EV1 so I can experiment with PCIe to Efinix FPGAs npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Well this is cursed. It's also exactly what I'm looking for https://www.amazon.com/Mini-Express-Extension-Adapter-Riser/dp/B01FVPITN8 npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg One thing I would like to do is get visibility into consumption of the HVAC unit. Does anyone know a good way to do that? It's fed by a 240V 30A circuit (2 hot + neutral); I have easy access to the power feed via splices in a junction box. Is there a readily available, affordable COTS device I can plug in there to slap some CTs around the hots and get realtime consumption data over SNMP or something? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg And which is more energy efficient if I'm not in a hurry. Am i better off using more cores for less time then idling, or just a few for a long time? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg So apparently running massive EM solver simulations makes my workstation draw enough power compared to normal usage that it shows up in the lab-wide power usage graphs. I wonder how much more it'll pull when I get the uncapped version that uses all my cores https://files.ioc.exchange/iocexchange/media_attachments/files/116/785/667/017/645/116/original/d091817c53c3c184.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…pfl7 Meanwhile, SPST... https://i.kym-cdn.com/entries/icons/original/000/036/050/Sans_titre_1.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Bit into a chocolate truffle from a local candy shop that was a kind of hemispherical shape, perhaps made from a flat chocolate base that had a blob of filling put on it and then encapsulated in more chocolate. And as I'm looking at the seam between the two segments the first thing that comes to mind is... Wouldn't it be cool to freeze one of these in LN2, cross section, polish, maybe ion mill, then do like SEM/EBSD to look at the cocoa butter grain structure around the weld line? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…5yze Sadly I don't get to dictate what products the school first aid kit stocks lol. If we find one she's not sensitive to, the school says we can bring a pack by and they'll designate it to be used instead of general inventory for any future minor injuries. I guess I'll run a few tests and see which styles/brands cause issues and which don't. But not until she's recovered from today, she was so excited to play with the neighbor's kid that she tripped over her own feet running out the front door and scraped up both elbows and a knee and had to go back inside to get patched up before playing lol. I ended up using 2x2 gauze and coban since there was no adhesive at all so I figured it was a pretty conservative approach as far as allergens go. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…5yze Sadly neither exists in the US to my knowledge, although the parent company may well sell the same product here under a different brand. The one that caused the issue was an amazonbasics product at school labeled "latex free" although with amazonbasics who knows how accurate that is lol. Couldn't find any obvious labels on the box as to if the adhesive was acrylic based or silicone based or whatever. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Do all bandaids use the same type of adhesive or are there multiple kinds in use? Kid seems to be allergic to at least some of them. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…luht Sounds like people aren't waiting long enough for the fire to get started before cooking because they're in a hurry. Getting a charcoal fire to the point it's really ready to cook can take like an hour depending on the conditions, which is more than enough to burn off any remaining lighter fluid residue. But if you start cooking at soon as the visible flames stop before it's fully ashed over there's likely still some residue. (Also you don't want to have the grill over the charcoal until it's almost ready to cook, unnecessary amounts of hydrocarbon soot on it probably won't taste great) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…a9g6 @nprofile…jckl (and paying a Dutch lawyer to answer them properly would almost certainly cost more than the training would pay) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…a9g6 @nprofile…jckl It would also depend on the specific legal definition of "education" - a formal accredited academic institution teaching courses in pursuit of a degree? Or any paid service whose primary function is to convey useful information to a student regardless of subject matter? These are not straightforward questions and I'm pretty sure the answers are all written in Dutch legalese npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…jckl Netherlands npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…kgg0 I care more about relative time/frequency than absolute time. I get out to around 11 significant digits with my setup which is... enough to get the job done npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…kgg0 (but if you really want good time depending on how crazy you are... GPS + PPS, PTP, or white rabbit) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Did you know that every catgirl has a maximum operating frequency determined by its size and dielectric constant? Exceeding the upper frequency limit can result in undesired secondary girlmodes rather than remaining in the intended TEM00. This can result in significant amounts of the catgirl's energy being wasted without any useful work being accomplished. You may be tempted to simply increase the catgirl's caffeination level to overcome the energy loss but modal dispersion makes this ill advised. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…kflw there's a reason I'm 99% WFH these days lol. I go into $dayjob to use their wire bonder, sem/fib, or meet with clients. Anything else i can do at home npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Which one of these BGA fanouts is better? Both are the same nominal trace/space, both are skew matched by the time you get to the vertical "home run" area heading towards the connector. But TDR of the first one wasn't that great. Let's see if the second one is any better when the simulation finishes... https://files.ioc.exchange/iocexchange/media_attachments/files/116/752/882/608/298/738/original/df02f2dece014b5e.png https://files.ioc.exchange/iocexchange/media_attachments/files/116/752/882/622/798/505/original/f90761a86b7556ae.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Looking at more rocks in the backyard with the kid. I think this is a conglomerate? The clasts are clearly harder than the matrix since they're sticking out while the matrix is eroding out from between them. Or is this more of a breccia? The clasts show some rounding but aren't fully rounded off. https://files.ioc.exchange/iocexchange/media_attachments/files/116/745/633/642/385/340/original/feac564248edb038.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg I don't know what kind of ceramic / glaze Thorlabs makes their mugs out of but when I'm microwaving a hot drink in one the handle is completely room temperature coming out. Almost no RF absorption or conduction from the hot liquid. Wonder what's different about them than, say, the Teledyne LeCroy mug where the handle gets quite toasty under the same conditions. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…3y57 the speeds and feeds are all wrong on every chocolate chip I've ever seen. Just look at how they're shaped, are they work hardening the stock or something?? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Just realized I forgot to include context... this is one of the fanout traces going from the XCKU5P on the LATENTRED switch mainboard to one of the Samtec ARF6 connectors that go to the QSGMII line cards and SFP28 uplinks. This particular trace is going to line card 0, PHY 0, lane 2 but is fairly representative of most of the others routed on the top side of the PCB (the ones coming in from the back with vias are separate and I'll do those next). I did some simulations a while back but I want to run through a whole suite now as I do final design review. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Ok so... First simulation pass on a SERDES track is done. Performance actually looks fine without a ground plane cutout under the BGA lands. https://files.ioc.exchange/iocexchange/media_attachments/files/116/734/083/617/183/460/original/ea72db5db522e456.png https://files.ioc.exchange/iocexchange/media_attachments/files/116/734/083/647/734/608/original/3ee92be02d1510a1.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…kuux white rabbit stereo system npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Found the first bug before even starting the SI simulations I had planned to be running. SFP28_0_TX is missing the ground plane voids under the BGA lands. But before I add them back it'll be a good excuse to simulate both launches and see how much a difference it really makes. https://files.ioc.exchange/iocexchange/media_attachments/files/116/729/681/664/908/656/original/5b820142852f57d3.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…k7h5 Once you finish the reballing tests if you send me the intact chip I wanted to test some higher end techniques on the mill for better-planarity substrate delayering. I can also do really high res imaging of each layer. Of course, if you want to try your hand at it, that's fine too npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Metal 1, coming up on the underside of the die. You can see all of the little wire bond rings, 4-5 concentric rings of wire bonds. I sanded up to the silver epoxy but not into the silicon to avoid potentially cracking it. So we should get a really clean die back side at the end of this. https://files.ioc.exchange/iocexchange/media_attachments/files/116/729/396/818/370/769/original/7895f53214148b05.jpg https://files.ioc.exchange/iocexchange/media_attachments/files/116/729/396/887/849/658/original/6ac27618a01edde4.jpg https://files.ioc.exchange/iocexchange/media_attachments/files/116/729/417/452/302/293/original/802b925c20eb1e77.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Metal 2. This turned out really good, planarity actually *improved* as I got deeper which surprised me because usually it's the opposite. Not sure if I was getting the feel of the part and correcting as I saw unevenness, or something about the way it was flexing, or what. But it seems like there is less edge rounding now. There's a few obvious thicker zone fills, maybe power tracks? And a lot of signal routes, including some length-matched diff pairs. Most of this layer is one big zone fill that I suspect is ground but have not probed to confirm. https://files.ioc.exchange/iocexchange/media_attachments/files/116/729/340/044/389/483/original/157b2459e271f7b3.jpg https://files.ioc.exchange/iocexchange/media_attachments/files/116/729/354/101/792/730/original/7bc772a41e5d40d2.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…k7h5 I haven't checked against the ftg256 pinout to see what they're matching but i found it interesting too. I wonder if they're trying to reduce overall package skew for dram ifaces? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg I could have stopped to image as the ring of copper shrunk inward, but I had good enough views through the soldermask that I didn't bother. Here's the next pass, with metal 4 fully uncovered and the edge of metal 3 polished away. We can see metal 2 and even 1 coming into view around the very corners. I want to experiment with using the mill for substrate thinning and see if this might avoid the edge rounding issue (assuming the substrate itself isn't warped). I'd likely still have to sand the last little bit, but it might be an improvement? I'll run a test of that later. https://files.ioc.exchange/iocexchange/media_attachments/files/116/728/595/130/870/377/original/d402cc4de1daebfa.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…7q7j "a 5 gallon Coke, got it. Will that be all? We offer a 20% combo discount if you get the 55 gallon drum of popcorn with it" npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…7q7j somebody's been looking at movie theater beverage sizes npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Hey @nprofile…a2ht new cursed idea if you're interested: QSFP+ to 4x RJ45 Ethernet module. I'm not sure if 4 rows, 4 columns, or 2x2 would look more absurd. As far as i can tell it's technically possible but doesn't exist, i hope it never is created, but I'd love to see just how derpy it would look lol npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…huwm I've put a ceiling of about 1.5 on my PUE measuring lab UPS vs whole building power draw. Don't yet have finer grained measurements to pin it down more precisely npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…9ylh for what sort of work? Tablet plus external keyboard is pretty portable and light as long as you're not doing anything too heavy. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg That's somewhat approximate because I don't know how accurate the 330W baseline number is (I've seen as low as 309W during idle-ish conditions) but I'm confident in saying each build costs me <2 cents in electricity. https://files.ioc.exchange/iocexchange/media_attachments/files/116/701/373/710/771/306/original/4493c9f9617b0dd6.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Just finished the first CI build with real time energy usage tracking. Peak consumption was 462W which is about a 130W jump over the baseline, but it doesn't stay that high the whole time. Total build time start to finish was 33 minutes, and total energy consumed (resetting the counter right about as the build began) was 254 Wh. Subtracting the baseline/idle power (330W * 33 minutes = 181 Wh) we get an estimated 73 Wh of delta power attributable to the build, not counting the contribution from the mac mini running the macos and aarch64 Debian builds which is probably small in comparison. At my current power rate of 0.218 USD/kWh after taxes and fees, and ignoring contributions from solar, this comes out to 1.6 cents of power used per CI build. Not bad. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg And looks like my normal baseline is ~330W with all of my normal base VM load but not pushing it too hard (i.e. not doing a CI build or anything) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg 270-277W at xcp-ng idle with no vms running npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Now to see how power hungry it is. 200-270W during POST lol npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg All done. Still more room for a single slot AMD card in the future. https://files.ioc.exchange/iocexchange/media_attachments/files/116/701/073/171/184/475/original/7bf8f88d6a7f3caf.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Soo according to Anthropic (https://www.anthropic.com/institute/recursive-self-improvement) they are thinking of slowing or halting frontier model development, but only if their competition does so as well and they have a nuclear-disarmament-style inspection process to verify this is the case. There's two possible reads on this, of course. The cover story is that they are worried about the models self-improving and becoming God and running away etc. Because they're vibe-coding 80% of their new stuff and model collapse is totally not a real thing. My preferred interpretation is that training costs are getting out of hand, they are terrified of shareholders finding out how much they are spending on training when they go public, so they want to cut back but they're afraid of falling behind and losing their edge so the only way they're willing to do so is if they can be assured the competition isn't out-spending them. There has to be an antitrust case here if somebody is willing to bring one (good luck in today's political climate but hey I can dream) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…k7h5 so how do usb a-c cables where there no cc on the far side work? Is there a resistor in the cable? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Cursed USB problem: I have a USB-C charge cable next to my bed that only charges my phone in one orientation. Flip it over and nothing happens. What single contact failure could cause this? Vbus and grouhd are redundant and it charges fine the other way around. If one of the CC pins failed open but vbus is present it should fall back to 500 mA 5V legacy usb slow charging right? npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Looking at our demand graphs and instantaneous usage, the hot water heater is actually one of our biggest non-lab power users (the power draw is pretty flat with big spikes every time we run the water heater). Our normal base house power usage is on the order of 5.5-6 kW but when the water heater is busy, it seems to spike to 10ish. Replacing the water heater has been on the TODO for a while due to it being old and probably rusting (plus not having any insulation between the tank and the exterior wall due to poor design of the utility room, which I plan to fix during the replacement process). Might go for a heat pump unit as a replacement. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Yesterday was another interesting solar record: the 88.6 kWh of solar production didn't quite beat Saturday's record of 93... but our grid power consumption was the lowest I can recall seeing. We consumed only 111 kWh, compared to the 150ish we more typically burn. Not entirely sure why... we didn't do any laundry or much in the way of cooking and it was a nice day so we had the windows open and no heat or A/C usage, which probably was part of it. Lab power demand was about the same as usual But combining the low consumption with the high production meant a net consumption of only 22.4 kWh, with a whopping 80% of our total power demand self-generated. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg I guess add this to the list of other things I wish the tech industry would build that they aren't... * Oscilloscope with nvlink or 100GbE * Laptop with SFP+ / QSFP+ NIC * Desktop GPU with 64+ GB of VRAM that doesn't need a small nuclear plant to power it. Rip out some of the tensor cores and put on more ram instead * ECC RAM as standard on desktop/laptop/mobile platforms * Cars with no ADAS or telematics And so on npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Sooo any chance the rtx spark will get to keep the big mellanox nic the dgx version has? A laptop with 100/200GbE would be awesome lol. (Looks at laptop with overpriced 10g thunderbolt dongle hanging off it) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…ec5a @nprofile…92fk @nprofile…m7ah @nprofile…m5ua @nprofile…vnlz maybe my business pipe is qos'd higher but i can hit the specced speeds with some consistency. Especially upload, during offsite backup refreshes i peg it for hours npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…92fk @nprofile…m7ah @nprofile…m5ua @nprofile…vnlz @nprofile…ec5a yep same here i can upgrade to 1g or more down but that's like 1g x 35 or something ludicrously unbalanced. I don't need more down i need more up npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…m7ah @nprofile…vnlz @nprofile…ec5a i want someone who will just run 10gbase-lr to my house and bill at 95th percentile usage or something but nobody does that :p npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…ec5a *looks at comcast business pipe with 35 Mbps upload and cries* npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…69vm @nprofile…2fsd rofl you have to do that on the qa stickers npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…2fsd i see your QA inspector is busy hanging out on the shelf while you do all the work npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Is anybody using, or considering using, ngscopeclient on an ARM64 platform? Would Debian aarch64 be a platform worth adding to CI and officially supporting? I think I can probably run it as a VM on the mac mini. But I'm not sure what's out there that actually has the GPU horsepower to perform well. Like, I think somebody tried to build it on a raspberry pi and it *worked* but I'm not sure how useful it would be on something so limited. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…29fk @nprofile…x8t7 yeah but like the other party won't accept it unless they can trade it on further npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…jstz The hard PCIe IP is from Cadence so I'm guessing this is probably like a Cadence PCIe gen4/5 PHY IP that they have some custom mux logic hooked up to so you can either wire up the PIPE interface to the PCIe hard block, or the raw SERDES to your gateware npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…jstz There is a DRP and it's standard AMBA APB although I haven't tried using it yet. Only configured through the GUI to date (the transceivers are not yet supported for RTL primitive instantiation although I am pushing them to add that in the future) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Tweaked the equalizer slightly and now it's passing the mask with no de-embed through the breakout fixture and cables https://files.ioc.exchange/iocexchange/media_attachments/files/116/656/849/543/998/169/original/77d9dcfb237b4747.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Yeah it looks like the de-embed opens the eye up by 30+ mV in the vertical axis but adds ~3ps of jitter which is just enough to push it over the edge into failing the mask test npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Huh, I'm actually getting better jitter performance with just a direct measurement and no de-embeds. I wonder if the S-params I'm using aren't as good as I thought they were npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg So far I haven't managed to get it to pass the SFF-8431 host eye mask with any equalizer settings I've tested (after de-embedding the SFP+ host compliance test fixture and the cable, so should be at the SFP+ connector mating plane). It's close, but there's a touch of overshoot I'm trying to eliminate plus there's a little too much jitter in the horizontal axis. https://files.ioc.exchange/iocexchange/media_attachments/files/116/656/718/622/283/630/original/9dcb20bf7b6d7824.png https://files.ioc.exchange/iocexchange/media_attachments/files/116/656/726/561/164/242/original/26038bfe192bd7fc.png npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Playing with the Efinix Ti375 SERDES https://files.ioc.exchange/iocexchange/media_attachments/files/116/656/688/602/928/450/original/3df3119551cdf7a3.jpg npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…tm9j yeah something like that. I'm sure there is a way to re-key i just have to figure it out npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg Starting to set up a second of these FS 40/100G switches. Both of them have the same SSH host key. That's... a problem lol. I need to figure out how to regenerate that... npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…t2th @nprofile…0f4l I am seriously looking at energy harvesting heat pumps to heat the house with waste heat from the lab in the winter, right now I run the air conditioner in January and feel like I'm wasting a ton of energy doing so. Because I am. npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…0f4l My actual server/network load is * 4U VM server that's my main big iron (probably pulls as much power as the rest of the rack) * Friend's 2U offsite backup server * 3x 2U Ceph nodes * 2U 100G router * 1U 10/40G Cisco Nexus switch I plan to retire soon, my former core * 1/2 width 1U FS 40/100G core switch that is going to get paired side by side with a second one later this week, at which point the Nexus should be redundant * 4x 1.5U Catalyst 24-port 1G switches, to be replaced with my own LATENTRED design one day (tm) npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…0f4l Wooow. Is that like actual equipment or do you have a ton of shelves and blanks and cable management bars and stuff? Let me see... * Left T&M rack: 14U of meat * Right T&M rack: 9U of meat * Main server/switch rack: 14U of servers, 7U of switching, 1U of NTP server = 22U of meat * UPS / devkit rack: 12U of UPS, 2U of defunct server I never unmounted, 2U of switching, 3 rack shelves full of FPGA boards = ~18U of meat So all told if you jammed everything in tight with no space used for patching or blanking I'd only be using ~63U or 1.5 racks npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…0f4l 3 full height racks??? Like, if you stacked my test equipment racks on top of each other in a standard height cabinet I'd have around two and a half but that is very spaced out and not optimized for density in the slightest, and includes a lot of things like switches providing LAN access to all of my workstations and wall ports around the house etc npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…jstz you can work with it but you need to plan for it, not throw stuff down on a board and figure it out later like you usually can in xilinx land npub1cddglts94qutscms0qpmk87lel9m8xku7q0wr20u2th5fxvvunqqxz9vpd Andrew Zonenberg @nprofile…jstz the thing to keep in mind about efinix is it's a lot of off the shelf third party ip and less in house vertical integration. So less flexibility, more attention needed to pin out and board design. Not all clocks can route to all PLLs, etc.