Tag
The author observes that used RTX 3090 GPUs are being sold on eBay for $1,300-$1,500, higher than a new 3090 Ti purchased 5 years ago, and questions why people buy older used GPUs at such high prices for AI rigs.
A tweet recommending users buy an RTX 3090 to run AI models locally, claiming it is easy, performant, and affordable.
A user details their modding and benchmarking of an AMD Strix Halo system with dual RTX 3090 eGPUs and NVLink, finding improvements in LLM inference speed for dense models, especially with vLLM, and discusses power efficiency trade-offs.
A reminder that two RTX 3090s and open-source models like Qwen 3.6 27B or Gemma 4 31B can run powerful local AI agents, comparable to Opus 4.5, using tools like Claude Code and self-hosted SearXNG.
Benchmarking the b9200 update of llama.cpp with optimized flags for Qwen 3.6 27B MTP on a single RTX 3090 shows significant performance gains, especially in prompt processing speed, for agentic workflows.
A user benchmarks the MTP variant of Qwen3.6 27B against the normal version on a single RTX 3090 using llama.cpp, finding MTP offers up to 2.37x faster generation at long contexts (32k-64k) but with slower prefill and no concurrency support yet.
A user shares power limit testing on a 4x RTX 3090 setup running Qwen3.6-27B with vLLM, finding 220W as the sweet spot for peak efficiency with minimal throughput loss.
PFlash now supports running @poolsideai's Laguna-XS.2 (33B-A3B MoE) on a single RTX 3090, achieving 111 tok/s decode and 5.4x faster prefill than llama.cpp, with NIAH passes up to 131K context.
The article presents benchmark results for 8 local LLMs on an RTX 3090, showing that power efficiency peaks around 225W, with diminishing returns at maximum power.