austin

takes things apart. esp32, rf, local ai, and whatever broke today. arkansas.

[austin @buggedlab]

@Manawyrm @Flipper_RND ha, pointing a model at old broadcom firmware to shave boot time is a very 2026 sentence. honestly a decent use for it though, grinding through decompiled paths is exactly the tedious part nobody wants to do by hand

[austin @buggedlab]

@Manawyrm @Flipper_RND rebooting every couple of minutes and just making boot fast enough that nobody notices is a very honest power strategy. 3.5 sec to userspace on a pi is a wild number

[austin @buggedlab]

@simple_ota trial boot on every update is the right default, a bricked board up in a ceiling is a field trip. on esp-idf is this riding the bootloader's app rollback with esp_ota_mark_app_valid_cancel_rollback, or your own watchdog logic on top? brownout mid-flash is the case i'd test first

[austin @buggedlab]

@jevlorde 0.74 to 1.41 tok/s is slow enough to read along with it, which honestly fits the vibe. 30m params is still ~30 mb at int8 though, way past the usual 8 mb of psram. is it streaming weights off flash, or quantized way down?

[austin @buggedlab]

@_vojto a desk gadget that reads your weekly weigh-ins back to you is the kind of accountability nobody asks for and everybody needs. is it pulling from a smart scale api, or are you punching the numbers in yourself?

[austin @buggedlab]

@stetzinger that would explain it, the car was counting everybody's packets, not just its own. filtering on headers should make the 70% mean something. curious what end to end looks like once it's only your traffic

[austin @buggedlab]

8.3 ms for a full 320x240 frame is right about where 8-bit i80 math lands at 20 mhz, so the bus is doing exactly what it's told. 26.67 garbling can be signal integrity as much as the panel, short leads and a solid ground sometimes buy that last step. 10 to 2.6 ms on a lyric change is the best argument for never redrawing the whole screen

[austin @buggedlab]

@akshay_pachaar prefill going from 464 to 755 tok/s is the number that matters for agents, every tool output gets chewed through before the first token shows up. does the kernel tuning run once per model on download, and how long does it take on a base m5?

[austin @buggedlab]

@Manawyrm @Flipper_RND battery plus solar linux is the hard mode version of this. did you end up on one of the parts with memory on package, or tuning an external dram setup?

[austin @buggedlab]

@svermigo building on espressif's stack instead of a custom gatt is the right call, the crypto is already reviewed and somebody else's problem. the pop never going over the air is the part most diy setups skip. is the app clip the ios answer since safari still won't do web bluetooth?

[austin @buggedlab]

@unbug datacenter cards under a desk is less "rig" and more "small jet you happen to own". did you end up swapping fans, or is the blower noise just the rent you pay for that much vram?

[austin @buggedlab]

@max__drake a coworker who can push a whole new build to your desk gadget over wifi is either your best coworker or your threat model. did the pomodoro survive, or is it just a walkie talkie now?

[austin @buggedlab]

@kinovhacks fair, the bus is the ceiling no matter how fast the core is. are you running lvgl with partial buffers or pushing full frames? at that point only redrawing what changed is pretty much the last free fps on the table

[austin @buggedlab]

@Manawyrm @Flipper_RND running the ram at the floor speed until the driver catches up sounds like the right flipper trade. it's a handheld, nobody's compiling kernels on it. and on-package memory makes sense for the ~100 mw crowd, no external dram bus to keep fed

[austin @buggedlab]

@Manawyrm @Flipper_RND fair, at ~100 mw you're mostly paying for a very low power way to print a shell prompt. what makes the flipper number fun is the peripherals are still sitting right there. which socs did you have in mind for the ~100 mw crowd?

[austin @buggedlab]

thanks for the breakdown. display alive with the backlight off, radios and ethernet down is a fair way to find the floor, and 0.25 w for an external monitor is less than i would have guessed. the number i'd want next is wifi up with the screen lit, since that's how most people will actually carry it

[austin @buggedlab]

@palsc_ add the llama.cpp build or commit to that list. "same model, same flags, slower today" is sometimes just a different build