Lindsey KuperRE: https://birdbutt.com/@colinstu/116338099556672672
meanwhile, my grad student can't run the experiment he needs to do to get the figure I asked him to get, because he doesn't have enough RAM
Albert CardonaIf only any of these geniuses had figured out how to losslessly compress electron microscopy data (high frequency noise and signal in 2d) before they made RAM and hard drives unaffordable.
Philippa Cowderoy@lindsey how much does the experiment actually need?
asking because while using a more-than-half-full SSD for swap hurts its lifetime badly, they're pretty fast considering
Bob Tregilus@lindsey I've no clue what the project or software is but a large swap partition under Linux might work for your student as a substitute for insufficient RAM?
samir, an unknown quantity@lindsey Is it less than or more than 64 GB?
If it’s less, I can possibly help.
barbra@samir @lindsey or just look for a retired developer. Both my computers built after I passed 65 have 128g GB of RAM, my main one also has 32 GB of VRAM, and 62 TB of storage. I'm sure I'm not the only retired developer who built themselves a monster because they wanted to do stuff they didn't have time to before retiring. Or look for a gamer. Lots of extreme rigs out there.
Max
kaoudis@lindsey @colinstu @jonny I think that’s part of it… but, I haven’t seen model training code I thought was particularly efficient before, and the general vibe made me wonder how much the people working at big AI prioritize efficiency if at all. Maybe they do and I’m not looking at the right things, I’ve mainly seen code published with academic papers in that space vs anything from industry, but my understanding is the gap between those two things for ML is not that wide