NVIDIA GeForce RTX 3090 — Benchmarks & Specs

Bottom line: how fast is the NVIDIA GeForce RTX 3090?

SpecPicks median: 93.8 tokens/sec across 6 cited single-user runs of 7-9B dense models at 4-bit (Q4) under llama.cpp (range 85.0–115 tok/s), from Hardware Corner (2 runs), GitHub: XiongjieDai/GPU-Benchmarks-on-LLM-Inference, LocalScore.ai, kunalganglani.com LLM Benchmarks, MyAIHardware. At the $1,999.99 Amazon price of the cheapest linked listing (checked 2026-10-21 12:20 UTC; prices change), that is a SpecPicks Value Index of 46.9 tok/s per $1,000. At 1440p Ultra, the NVIDIA GeForce RTX 3090 averages 92 fps in Tom's Hardware rasterization test suite (geometric mean), per Tom's Hardware. For local LLM inference it generates 135.7 tokens/sec running qwen3.6:35b-a3b (a mixture-of-experts model with ~3B active parameters, so dense models of similar size run far slower) at Q4_K_XL under llama.cpp, per GitHub: thc1006/qwen3.6-speculative-decoding-rtx3090. In PassMark G3D Mark it scores 26,546–26,671 pts, across three published results (PCBench, PassMark GPU benchmark listing, PassMark GPU benchmark listing). Its 24 GB of VRAM is the binding constraint for local inference: the qwen3.6:35b-a3b run above (35B parameters) is the largest model on file at 4-bit quantization on this card. That run used a 16,384-token context; the Will-It-Run matrix sizes each model at its listed quantization with an ~8K context, and by that measure DeepSeek-R1-Distill-Qwen 32B at Q4_K_M is the largest that fits in 24 GB.

Every figure above is a row in the tables below, and each row links out to the review or public benchmark database the number was taken from; the SpecPicks median is computed here from the runs it names, each a row in the AI table, and the per-$1,000 figure divides it by the dated listing price. SpecPicks aggregates published measurements; it does not report first-party benchmark runs.