NVIDIA RTX 5090 vs NVIDIA RTX 3090

Analysis

The RTX 5090 and RTX 3090 sit five years and two architecture generations apart (Ampere to Blackwell), and at first glance this looks like a pure speed upgrade: newer flagship against the used-market card it eventually replaced. The comparison is more interesting than that once VRAM capacity, not just bandwidth, actually enters the picture.

The RTX 5090's 1,792 GB/s is 91.5% faster than the RTX 3090's 936 GB/s, and that gap shows up almost exactly in tokens per second on any model both cards fit: Llama 3.1 8B at Q4_K_M decodes at 195.9 tok/s on the RTX 5090 versus 102.3 tok/s on the RTX 3090, and GPT-OSS 20B at Q4_K_M runs 155.0 tok/s versus 81.0 tok/s, both tracking that same 91.5% ratio closely. The RTX 5090's extra 8GB of VRAM (32GB versus 24GB) also changes what fits, not just how fast: Qwen3 32B at Q4_K_M runs natively on the RTX 5090 at 54.6 tok/s (23.88 GB, comfortably inside its roughly 30.4 GB usable budget), while the identical build needs CPU offload on the RTX 3090's smaller roughly 22.8 GB budget, dropping to 28.5 tok/s. Across this site's full tracked catalog, that extra capacity only widens the native-fit list by one model (53 of 94 versus 52 of 94), since most of the gap between 24GB and 32GB still isn't enough to cross the next real threshold: 70B-class dense models. That's the twist in this comparison: even the RTX 5090, at nearly double the price of a used RTX 3090 and with a real generational bandwidth and VRAM advantage, still can't run Llama 3.3 70B natively at 8k context on its own; its roughly 30.4 GB usable budget falls short of the 40.72 GB a Q3_K_M build needs. Two RTX 3090s pooled over NVLink (or plain PCIe layer-splitting) reach that same Q3_K_M build natively at 15.1 tok/s, something no single card in this comparison, RTX 5090 included, can do alone.

Bottom line: For anyone buying one card, the RTX 5090 wins outright: faster on everything the RTX 3090 already runs, and its extra VRAM opens a little more headroom at the 32B tier. But 'one card' is the operative constraint, and it's worth being explicit about what it doesn't buy: 70B-class local inference, which still needs either a bigger workstation card or a pair of pooled 24GB cards. A used RTX 3090 pair costs meaningfully less than one RTX 5090 and gets there; a single RTX 5090 does not. Choose the RTX 5090 for the fastest single-card experience on everything up to the 30B class; choose two RTX 3090s specifically if 70B-class models are the goal and budget matters more than fitting everything on one card.