NVIDIA GeForce RTX 5090, 32GB vs NVIDIA GeForce RTX 4080, 16GB for local AI
The GeForce RTX 5090, 32GB holds 34 of the 50 current models here, and the GeForce RTX 4080, 16GB holds 16. The GeForce RTX 4080, 16GB costs $800 less, both priced as the card alone, without the PC around either. On Gemma 4 12B, the strongest model both hold, the GeForce RTX 5090, 32GB is about 2.1× faster: 155 tok/s against 74, both estimated from memory bandwidth. The GeForce RTX 4080, 16GB pays for itself sooner, in 93 years against 123 years at 500k tokens a day, though that is each machine on its own strongest model rather than on the same one. The GeForce RTX 4080, 16GB is the previous generation, so every figure here for it is priced at what it launched at rather than at a price you can pay today.
| NVIDIA GeForce RTX 5090, 32GB | NVIDIA GeForce RTX 4080, 16GB | |
|---|---|---|
| Price | $1,999card only | $1,199card only |
| Memory | 32 GB | 16 GB |
| Usable by the GPU | 31 GB | 15 GB |
| Memory bandwidth | 1792 GB/s | 716.8 GB/s |
| Power under load | 575 W | 320 W |
| Models that fit | 34 | 16 |
| Best model it runs | Qwen3.8 27B | Gemma 4 12B |
| Speed on that model | 66 tok/s estimatedQwen3.8 27B | 74 tok/s estimatedGemma 4 12B |
| Pay-back on that model | Pays back in 123 years | Pays back in 93 years |
Run the numbers on the NVIDIA GeForce RTX 5090, 32GB · or the NVIDIA GeForce RTX 4080, 16GB
Jump to: Side by side on Gemma 4 12B · How much use it takes to pay back · What the extra memory buys · The assumptions behind both columns
Side by side on Gemma 4 12B
The table above gives each machine the strongest model it can hold, and those are not the same model, so the two speeds in it are not a race. Gemma 4 12B is the strongest model both machines hold, so this is the pair running the same work.
| NVIDIA GeForce RTX 5090, 32GB | NVIDIA GeForce RTX 4080, 16GB | |
|---|---|---|
| Speed | 155 tok/s estimated | 74 tok/s estimated |
| Pay-back | Pays back in 152 years | Pays back in 93 years |
Run Gemma 4 12B on the NVIDIA GeForce RTX 5090, 32GB · or on the NVIDIA GeForce RTX 4080, 16GB
How much use it takes to pay back
Everything above is at 500k tokens a day. Pay-back moves with how much you actually run, so here are both machines on the same model, at the five levels of use the calculator names. The NVIDIA GeForce RTX 4080, 16GB pays back sooner at every level of use, so this is not a choice that turns on how hard you work it.
| A day's use | NVIDIA GeForce RTX 5090, 32GB | NVIDIA GeForce RTX 4080, 16GB |
|---|---|---|
| 50ka few chats a day | 1,517 years | 934 years |
| 200klight assistant use | 379 years | 234 years |
| 1Ma moderate coding-assistant day | 76 years | 47 years |
| 4Mheavy coding with an agent | 19 years | 12 years |
| 20Magents running most of the day | 3.8 years | 2.3 years |
Run the NVIDIA GeForce RTX 5090, 32GB at 20M tokens a day · or the NVIDIA GeForce RTX 4080, 16GB
What the extra memory buys
The NVIDIA GeForce RTX 5090, 32GB holds 18 models the NVIDIA GeForce RTX 4080, 16GB cannot at 32k of context. The strongest of them are what the difference in memory actually buys.
| Model | Weights | Needs at 32k | On the NVIDIA GeForce RTX 5090, 32GB |
|---|---|---|---|
| Qwen3.8 27BSonnet-class | 16 GB | 19 GB | 66 tok/s estimated |
| Qwen3.6 27BHaiku-class | 17 GB | 19 GB | 65 tok/s estimated |
| Nemotron 3.5 Bolt 74B-A9BHaiku-class | 22 GB | 22 GB | 203 tok/s estimated |
| Qwen3.6 35B-A3BHaiku-class | 22 GB | 23 GB | 235 tok/s estimated |
| Muse Glimmer 30BHaiku-class | 17 GB | 18 GB | 69 tok/s estimated |
| Gemma 4 26B-A4BHaiku-class | 17 GB | 18 GB | 172 tok/s estimated |
12 more, on the NVIDIA GeForce RTX 5090, 32GB page.
The extra memory buys context as well. Of the 16 current models both machines hold at 32k, 4 run to a longer window on the NVIDIA GeForce RTX 5090, 32GB: the weights are a fixed size and the KV cache is not, so what the weights leave spare is what a longer context grows into. Granite 4.2 8B reaches 128k there against 32k on the NVIDIA GeForce RTX 4080, 16GB, each the longest window the calculator offers that the machine still holds it at.
The GeForce RTX 5090, 32GB is also head to head with another card: RTX PRO 6000 Blackwell, 96GB · GeForce RTX 4090, 24GB · GeForce RTX 3090, 24GB · Radeon AI PRO R9700, 32GB · GeForce RTX 3060, 12GB. With a complete computer: Framework Desktop, 64GB · MacBook Pro M5 (14-inch), 16GB.
The GeForce RTX 4080, 16GB is also head to head with another card: RTX PRO 6000 Blackwell, 96GB · GeForce RTX 4090, 24GB · GeForce RTX 3090, 24GB · Radeon AI PRO R9700, 32GB · GeForce RTX 3060, 12GB.
The assumptions behind both columns
Both columns use the same usage: 500k tokens a day at 15:1 input to output, 32k of context, $0.17 per kWh, and today's API prices held flat. Speeds marked estimated are worked out from memory bandwidth rather than measured, and pay-back scales with them. Both are priced as the card alone, so neither figure includes the PC to put it in. All seven cards here are ranked by what each one holds. Change any of it in the calculator.
More head to head: everything the NVIDIA GeForce RTX 5090, 32GB runs · everything the NVIDIA GeForce RTX 4080, 16GB runs · every other match-up · the quickest pay-back at each level of use · every model against the frontier