Before plugging both cards in together, I checked the new card's stability on its own — and dug up an old idea I'd shelved as "no effect" and flipped it
This is the English version of a post originally written in Korean for my algorithmic trading system devlog(new tab).
A few days ago I'd already decided on the direction for a hardware expansion, and today I actually carried it out. Along the way I also dug up an old idea I'd given up on and flipped it.
GPU expansion: swapped in the new card first to test it alone
I picked up a second graphics card, the same model as the one already running. The end goal is to run both at once, but today I just swapped in the new card and tested its stability solo first.
I ran a load test with power slightly capped. Temperatures actually stayed lower than before, fan noise dropped, and there were no slowdowns or errors at all.
I haven't actually installed both cards together yet. Checking the case's internal layout ahead of time, I found that the slot for two cards side by side collides with the power supply's mounting spot by just a few centimeters.
Checking the dimensions before blindly assembling meant I could reach the conclusion — "I need a case that puts the power supply in its own separate compartment" — before actually running into the collision. Running both cards together for real is on hold until that case swap happens.
I dug up something I'd shelved as "no effect" and flipped it
A while back I tried letting the local LLM server handle multiple requests concurrently, hoping it would speed things up. I measured it, concluded "no real effect," and shelved the setting.
Today I pulled it back out under different conditions and measured it properly, and the result was completely different this time — nearly double the raw generation speed, and about 1.5x on the full task end to end.
Thinking back on why the old conclusion was wrong, the measurement method or conditions back then were probably different enough that the improvement was masked by some other bottleneck. I learned again that "it didn't work before" has an expiration date once conditions change — when the premise shifts, an old idea shelved before is sometimes worth pulling back out.
Also today
- An experiment comparing two local LLMs (a lighter, faster one vs. a heavier, slower one) on how much they actually sway investment decisions cleared a major hurdle. It's taking longer than expected, so results are still a ways off.
- I'd mentioned moving the core processes to OS-level services a while back. Today I actually rebooted the machine and confirmed all four came back up on their own without any manual intervention. During the check I briefly chased a false alarm — I'd queried a service by the wrong name and thought "why is this down?" when it had been running fine the whole time.
Today leaned more toward hardware than software. Running both cards together is left for the next step, and today's attitude was: check what needs checking first, before jumping ahead.
Top comments (0)