Libre Computer

Systemd needs to make the only choice that is reasonable and not support systems in California, Colorado, and Brazil if laws pass.

Feel like it is only a matter of time before DRAM manufacturers become compute manufacturers.

@bretweber I mean MCUs 😀

@librecomputer Still trying to stick a price point. I respect it

April will be amazing.

@bretweber 4MB of RAM is where it is at.

@librecomputer Libre Computer Green Bean releasing with 128MB of RAM?

Connecting with people in GPU sales/leasing/cloud compute, containerized/modular solutions, datacenter design & deployment space. Have 2MW+ cheap reliable power source for PoC, expandable to multiple sites.

@bretweber Barrel plug without flexible input is crazy.

How many of these do we think ended up cooked before they started putting these stickers on them?

The bottleneck will ping pong between wafer output, datacenter construction, and power generation/delivery for the foreseeable future.

@sahajsarup We don't so not every. Stock Debian and Fedora.

Why does every freaking sbc OS have a complicated "startup script"? Whatever happened to just dropping to console? Armbian/Rpi/DietPi whatever else l. Everyone wants to change the password, connect to network, update every darn thing and do full desktop configuration 🤬

@axboe @lffgz Do you ask it to write a bunch of md files guiding intuition? This is what we found most helpful through OpenClaw. Building context and groundwork of relative truths.

@lffgz I don't think you can, because that skill comes from intuition built from years/decades of experience. Same goes for taste, though many experienced developers never achieve that.

Experiment of the day - have Claude debug an issue, where I have a rough idea of what may be wrong. Here's the crash: syzkaller.appspot.com/x/log.txt?x=14… and here's what Claude spat out after chewing on this for a while: git.kernel.org/pub/scm/linux/… which I instantly knew was likely correct.

@axboe @erwanaliasr1 We have been uaing Claude with great success. For architects who understand the implications in the stack, it is great. For vibe coding, it is a recipe for disaster.

@erwanaliasr1 Claude wrote it.

UEFI extensibility? Device tree will always be superior.

@bretweber @Raspberry_Pi Yeah that is a good stress that almost all controllers optimize for because the benchmarks do that test.

Hm, well this 1st iteration was generating a deterministic pseudo-random data per file (each cycle of 95% of the disk's capacity gets a new file) using a seeded PRNG (Go's math/rand), bypassing the OS caches with O_DIRECT, and then reading it back to compare the SHA-256 hashes. I haven't deployed anything big time yet so I'm open to tweaking. This was a side quest that I haven't really dedicated much time to :D

Almost at 50TiB written/read on the 32GB @Raspberry_Pi microSD card, and no smoke yet.. I should really try and find time to better understand the innards of flash storage, the increase (and then decrease) is interesting. Controller optimising for the workload for a while maybe?

@bretweber @Raspberry_Pi The real test is writing 4K data randomly and explicitly sychronized. That will destroy poorly designed devices with only a few GB.

@bretweber @Raspberry_Pi NAND can take a lot of sequential writes. Take your size and multiply by 1000 should not be an issue. You have to write random data (eg AES encrypted via openssl) to really test as flash controller have compression now to improve perf and not actually write much.

@System64fumo 8GB is $100 BOM now and can no longer be considered low end. 🤣

@librecomputer Depends on what “good” and “low hardware means”, I feel like a local translation or automation bot is helpful and that can run on pretty low-ish hw (8gb ram) But generally speaking you’re correct, Actually useful coding assistants or similar do need beefy hw.

@librecomputer You guys have been promoting AI/Agent stuff lately, Would be super cool to see an LLM/Agent run fully locally running on upstream linux and llama.cpp with HW accel since rk3588 or similar can’t utilize the NPU/GPU with llama.cpp.

@System64fumo Local models are only good for roles such as TTS or STT or image classification. Useful LLMs cannot run effectively on lower end hardware.

@librecomputer A small low power cluster that can utilize an NPU or GPU to accelerate an LLM would be very appealing because at the moment the only way to get hardware accelerated LLMs on ARM is with vendor kernels and very specific and broken AI tools.

@geerlingguy needs to create Clawsible to replace Ansible.