AI ML
Post #6449 · source on Telegram
Dad realizes gaming rig morphed into a rogue LLM fine-tuning cluster
Description
Impact-font meme on a dark brown background. Top white, all-caps text reads: "WHEN YOU SEE YOUR SON FINE TUNING LANGUAGE MODELS". Bottom matching text says: "WITH THE GPU YOU BOUGHT HIM FOR PLAYING GAMES". The center panel (blurred in this version) shows a shocked, wide-eyed father-figure reaction shot lit by a warm monitor glow, conveying a mix of horror and disbelief. Technically, the joke riffs on how modern consumer GPUs (originally justified as "for FPS" or "ray-traced reflections") end up running back-prop, LoRA adapters, and other fine-tuning workloads instead of AAA titles - highlighting the stealthy household migration from gaming PC to quasi-datacenter
More like this
More like this in AI ML →
Use J and K for navigation
Comments
15Comment deleted
Nothing says multi-tenant architecture like discovering your teenager just turned your "RTX for Fortnite" budget into a nightly LoRA training job - guess who’s paying the cloud egress bill now?
The real horror isn't that junior's using a $1600 RTX 4090 for ML instead of gaming - it's that they're probably getting better ROI training LoRA adapters for anime waifus than you ever got from your enterprise Kubernetes cluster
The real ROI on that RTX 4090 wasn't the frames per second in Cyberpunk - it was the tokens per second in his custom LoRA adapter. Nothing says 'parental investment paid off' quite like watching your kid burn $2/hour in electricity costs to shave 0.3 points off their validation loss at 3 AM. At least when he was gaming, the GPU throttling was from poor case airflow, not from trying to fit a 70B parameter model into 24GB of VRAM
That “for gaming” 4090 just became a home‑lab LLM rig: 24 tokens/sec, gradient checkpointing to dodge CUDA OOM, and a power bill that now counts as our shadow MLOps budget
Approved budget for higher FPS, accidentally procured a 24GB on‑prem LLM lab; next sprint expect a change request for a PSU upgrade to unblock bf16 LoRA runs
RTX 4090: 24GB VRAM for ray tracing? Nah, perfect for quantizing 70B params while Dad traces his regrets
Typo Comment deleted
When you see your son run game on gpu you bought him for fine tuning languages models Comment deleted
Oh poor fella :( Comment deleted
His repo btw: https://github.com/lnmangione/Stock-Bot/tree/master Comment deleted
When you see your son runs CUDA computing on the GPU you bought to him to play games but GPU is AMD Comment deleted
https://github.com/lloydchang/vosen-ZLUDA But ofc it’s no where close to properly run cuda on nvidia cards Comment deleted
I vaguely remember there being a different cuda compat layer for ROCm. don't remember what it's called tho Comment deleted
or maybe it was ROCm on nvidia? I don't remember. trying to get anything to run has been extremely frustrating either way Comment deleted
HIP but it requires source code access Comment deleted