Hardware Corner RTX 3090 local LLM benchmark page (March 2026)

Per-SKU llama.cpp table, Updated Benchmarks: March 2026. Single RTX 3090 (24 GB GDDR6X, 936 GB/s as printed on the page). Tables cover Qwen3 8B/14B/32B Q4_K, gpt-oss 20B MXFP4 and Qwen3 30B-A3B Q4_K at 4k through the largest context that fit. Token-generation column is batch-1 per-request decode; prompt-processing column is prefill at that KV depth. The page does not publish the llama.cpp build hash, CUDA version, or exact prompt/output token counts beyond the context-depth bucket.