ToolScore

Local GPU AI

7 articles — every claim carries a date and a source.

What Actually Runs on 12GB: RTX 3060 vs 3080 Ti, MoE Speed + Max Context, Measured

128k context from 12GB

Every number measured live with llama.cpp b10872 on both cards in the same box: gpt-oss-20b hits 103–110 t/s, Qwen3-30B-A3B at Q2 runs 77–108 t/s mixed VRAM+RAM and stretches to a 128k context at q4_0 KV, and the 35B Qwen3.5/3.6 class lands 43–100 t/s. Full flag-by-flag configs, the context ceilings per model, and the CUDA VMM trap that fakes a successful launch and crashes on the first prompt.

12 min read

Used GPUs Over $800 for Local AI: The Cost-per-GB Collapse and the $3,500 Mining-Card Gamble

~$99 per GB of VRAM

Used-US sold medians, human-verified 2026-11-10 (online ask/BIN runs higher): above $800 each rung roughly doubles the cost per GB — ~$50 (3090), ~$99 (4090), ~$216 (5090). What each tier buys, why the value collapses, and the CMP 170HX unlock story with its 2-in-5 failure rate.

5 min read

Budget Local AI Under $800: Two Used RTX 3060s Get You 24GB of VRAM at ~$27 per GB

~$27 per GB of VRAM

Used-US sold medians, human-verified 2026-11-10 (online ask/BIN runs higher): the real sub-$800 local AI build is a pair of RTX 3060 12GB cards — 24GB total for $638, $420–500 local. What actually runs there, including a 35B MoE at 118–160 t/s, and the no-miracles limits of the 27B class.

6 min read

2× RTX 3060 vs 1× RTX 3090: Same 24GB of VRAM, ~$560 Apart

~$560 saved

Used-US sold medians, human-verified 2026-11-10 (online ask/BIN runs higher): a pair of RTX 3060 12GB ($638) and a single RTX 3090 ($1,199) both deliver 24GB for local AI. Where the pair wins (cost per GB, MoE serving), where the 3090 wins (one slot, ~80 t/s on 27B, image work), and what the $560 difference buys instead.

4 min read

GPU Rent vs Buy: The 5,700-Hour Math on a Used RTX 3090

~5,700 hours to break even

Renting a 3090 runs about $0.21/hour (vast.ai median) while used cards sell at $1,199–$6,919. The break-even math with electricity included for the 3090, 4090 and 5090 — and why the answer is always a hybrid.

5 min read

Rented GPU vs API: $0.27 Per Million Tokens — When Each One Actually Wins

~$0.27 per 1M output tokens

A rented 2x RTX 4090 pod serving 985 t/s costs about $0.27 per 1M output tokens at vast.ai median rates — the per-token math against OpenRouter pricing nobody has published, plus the honest privacy picture of timeshared GPUs.

5 min read

Unified Memory, Honest Numbers: 284B on a Laptop, 26 t/s, and the Bandwidth Ceiling

~26 t/s — 284B on a laptop

Apple M4 Max, M5 Ultra, DGX Spark, Strix Halo and the R9700 compared on verified speeds and prices — and why memory bandwidth, not GB, decides what unified memory can actually run.

6 min read