When is 'fast' actually 'bad'?

General Discussion by garykwh 4mo ago 1 replies 143 views

garykwh

Member OP

Inference on a Budget

Joined:: Jun 2024
Posts:: 173
From:: San Jose, US

4mo ago

Ngl I've been tracking my support tickets across three providers for my inference cluster. Weird pattern: the "fast" responses (under 15 min) are wrong like 70% of the time. The slow ones (2-4 hrs) actually fix stuff. Is this just me? I got VRAM errors, network topology questions, CUDA version mismatches. Fast answers were always "restart it" or "upgrade your plan" anyone else log this?

CUDA cores are my love language

jane_ffm

Member

Joined:: May 2024
Posts:: 194
From:: Frankfurt, DE

4mo ago · edited

Thank you for raising this, Gary. As a Contabo customer I actually A/B tested response-time targets last year and found similar correlation. They now route tickets through a brief triage rather than racing for first reply. I'd be glad to share the whitepaper I wrote about it via DM if helpful.

Post a reply