My office until 100k MRR
by•
Macbook air m5 : Building (i don't need more ressources because i'm building @Murmell with @Murmell ) adn posting across socials
Desktop (RTX 5090 + ryzen 7800x3D) : Some uncensored models for cybersecurity tasks + local Gitlab runner
Laptop Lenovo Legion 5 : Test environnement before deploying to the production and playing some Fifa lol
40 views
Replies
the air is the surprising one in that list. sustained inference on a fanless machine behaves completely differently from bursty work, it holds peak for a couple of minutes and then quietly halves on you, and that drop is very easy to misread as the model being slow rather than the chassis being hot. so the split that actually matters isnt mac versus pc, its which box can hold load and which one is fast in short bursts. sounds like youve already landed on that by accident with the 5090 taking the long jobs. the local gitlab runner is the bit id steal though. ci that doesnt bill you per minute quietly changes how often you let an agent retry something, and that turns out to matter more than raw speed.
Murmell
@rabnoor_s
Yeah, exactly 😄 I’ve noticed the same thing with the Air, it’s great for short bursts, but once you keep it under sustained load, the thermals start to matter.
That’s why I’ve been splitting the workloads between the machines rather than trying to make one machine do everything.
And yeah, the local GitLab runner has been surprisingly useful. Being able to let agents retry / run CI without worrying about per-minute costs changes the way you use them quite a bit.
@moss_ab_mirande_ney the free retries point is the one i keep coming back to. when a rerun costs nothing you stop writing for one correct attempt and start writing for cheap repeated ones, and that quietly changes what you even ask for. today i ran a verification pass i would have skipped if it were metered, and it caught something i was about to ship wrong.
Murmell
@rabnoor_s Yes of course, for me it's more on the delay question, for example, with github, a standard deployment take aboutt 10 - 15 minutes, with my runner it's about 20 to 40 seconds so it's way more efficient.
I'm confortable on the cost side, i had the chance to benefit of 50k minutes of Github actions due to the microsoft startup program.