RTX 4090 vs RTX 5090 for Local AI: What 8GB More Actually Changes

RTX 4090 vs RTX 5090: the specs side by side

Where the fit differs: 24GB vs 32GB

We ran 7 models across five quantization and context settings through the hardware calculator's formula on both cards. These are the only 6 combinations where the two cards give different answers. Everything else gets the same verdict on both.

The pattern: anything the 4090 handles comfortably, the 5090 handles too, and the 5090's gains cluster in the 24B to 32B range. If the models you run are 14B or smaller, both cards fit them with plenty of room, and the extra 8 GB buys you nothing you'll notice. The full per-card tables are on the RTX 4090 and RTX 5090 pages.

*Gemma 3 figures are on the cautious side. Gemma 3's published config keeps only a 1,024-token window for five of every six layers, but the calculator counts every layer at full context. Runtimes that store just that window for those layers need less memory than shown, especially at 32K.

Speed: bandwidth and compute

Once a model fits, how fast it generates text is mostly limited by memory bandwidth, and the 5090's is about 78% higher. It also has about a third more CUDA cores, which helps when the card is working through a long prompt before it answers. NVIDIA lists FP4 support for the 5090's Tensor Core generation, which matters only for software and models built to use it.

We haven't benchmarked either card, so this page doesn't put a tokens-per-second figure on the gap. The bandwidth numbers are the honest starting point: for the same model on both cards, expect the 5090 to be clearly faster, but not 78% faster in every case, since other parts of the system play a role too.

The 70B question: one 5090 or two 4090s?

Neither card runs Llama 3.3 70B at Q4 on its own. It needs about 49.1 GB by the calculator's math. That raises a common question: is one 5090 or a pair of 4090s the better route?

• Two 4090s give you 48 GB, which puts 70B at Q4 in tight territory at best, before the overhead of splitting a model across cards, which the calculator doesn't model. That's 900 W of graphics cards, and the 4090 has no NVLink, so the cards talk over PCIe. Ollama's docs say it spreads a model across all available GPUs when it doesn't fit on one.
• One 5090 gives you 32 GB on one card: no splitting, simpler power planning, and comfortable room for 32B models, but no 70B at Q4.

If 70B models are a hard requirement, neither option is comfortable. The Apple Silicon guide covers Macs with enough unified memory to hold them on one machine.

Power and running cost

The 5090 is rated 125 W higher, and NVIDIA's recommended power supply goes from 850 W to 1,000 W. Both use the 16-pin connector, but the 5090 wants a 600 W cable or a four 8-pin adapter instead of the 4090's three. If your current power supply was sized for a 4090, check it before you swap cards.

The calculator uses each card's rated power, so treat the results as an upper limit. A card generating text often draws less than its rating, and an idle card draws far less.

Which one to buy for local AI

Pick the RTX 5090 if you want 32B models like Qwen3 32B or DeepSeek-R1-Distill 32B running with room to spare, you want Q5 or Q8 on 24B to 32B models, or you want the fastest single consumer card for generation.

Pick the RTX 4090 if the models you run are 27B or smaller, you can find one for meaningfully less than a 5090, or your power supply and case were built around 24GB-class cards. If capacity is your only concern, also look at the RTX 3090, which holds the same 24 GB as the 4090.

Whichever you lean toward, compare real prices. The 4090 reportedly wound down production in late 2024, so most listings are used or renewed. The 5090's launch price was $1,999, and what it sells for can be a different number.