NVIDIA RTX 5090 vs NVIDIA RTX 4080

Analysis

The RTX 5090 and RTX 4080 sit two architecture generations and roughly $800 apart: Blackwell's $1,999 flagship against Ada Lovelace's $1,199 mid-tier card from 2022. The gap in local LLM capability is bigger than that roughly 67% price difference alone suggests, both from the sheer bandwidth and VRAM increase and because the RTX 5090 runs an entire quantization format the RTX 4080 can't.

The RTX 5090's 32GB of GDDR7 on a 512-bit bus delivers 1,792 GB/s: exactly 2x this card's VRAM and 2.50x its 717 GB/s bandwidth. Since decode is bandwidth-bound, that tracks almost exactly in this site's calculator: GPT-OSS 20B at its recommended Q4_K_M decodes at 155 tok/s on the RTX 5090 versus 62 tok/s on the RTX 4080, a 2.50x gap matching the bandwidth ratio precisely. The RTX 5090's 5th-gen Tensor Cores also add native NVFP4 support, a 4-bit format Ada Lovelace's architecture can't run at all in this site's calculator: Qwen3 32B fits the RTX 5090 natively at NVFP4 (19.87 GB, 65.7 tok/s), a model that doesn't fit the RTX 4080 at any quantization on the standard GGUF ladder, landing at just 9.1 tok/s offloaded at Q3_K_M there instead. Qwen 3.6 27B shows the same pattern at a smaller scale: NVFP4 lets it fit the RTX 5090 with real headroom (15.72 GB, 83 tok/s), while the RTX 4080 needs its more conservative Q3_K_M build (15.15 GB, 34.5 tok/s) just to stay native.

Bottom line: This isn't a close call: the RTX 5090 doubles VRAM, more than doubles bandwidth, and adds a quantization format the RTX 4080 simply can't use, for roughly 67% more money. For anyone whose budget stretches to $1,999, the RTX 5090 is the better card on every LLM-relevant axis. The RTX 4080 mainly makes sense as a secondhand or discounted buy for workloads that stay comfortably inside 16GB; anyone specifically targeting 27-32B-class dense models should budget for the RTX 5090, or a 24GB card like the RTX 4090, rather than expect the RTX 4080 to stretch there.