glm 5.3 and qwen 3.8 already run at good quants on the new rig. deepseek is the boss fight

not rich, just bought it one gpu at a time 😂

threadripper 3975wx on a gigabyte mc62-g40. 7 slots, one bifurcated so all 8 fit

🤣🤣🤣🤣🤣

haven't benchmarked tensor parallel on it yet. testing it properly soon and i'll share numbers

where to grab the main parts: board (mc62-g40): newegg.com/p/pl?d=mc62-g40 cpu (3975wx): amazon.com/dp/B08V5H7GPM add2psu: amazon.com/dp/B0794KMV4B 4tb sn5000: amazon.com/dp/B0D7MLB76V 3090s are all used, ebay is the move

day-vault parts list: - 8x RTX 3090 (3 Dell, 2 Zotac, EVGA, ASUS, FE), 160w cap - Threadripper PRO 3975WX (going in soon) - Gigabyte MC62-G40, one slot bifurcated - 224GB DDR4 (7x32, 8th slot dead lol) - 1600w + 850w PSU via add2psu - 4TB WD SN5000 + 512GB drive - new rack soon

bought it piece by piece, all at once would be closer to 30k. 160w cap per card for now. glm 5.3 and qwen 3.8 run great, tok/s numbers coming

cards are capped at 160w each, so it's nowhere near the 2-3kw people are guessing 😄

I could use a matcha

Interesting I may try this

dunk the lunk to the smartest guy in the mansion. hell of a glow up

under a second is where it stops feeling like a walkie talkie. the redo on a reply that breaks halfway matters just as much though, nothing kills a voice demo like it freezing mid sentence

I could use some for a good meal tomorrow

Yes

the uncapped 350w part matters. readme says 225w cards only got +25%, so my 160w-capped 3090s probably see a prefill bump at best