NV NVIDIA Consumer Discrete GPU

NVIDIA GeForce RTX 4090

Runs downloadable LLMs up to ~32.5B at Q4, ~164.3-195.8 tok/s on an 8B model.

NVIDIA GeForce RTX 4090
24GB
VRAM
1008
GB/s bandwidth
32.5B
Max model (Q4)
$1,599
MSRP

What you can run

At 8k context, FP16 KV. Full-speed tok/s holds up to "max ctx"; beyond it the context spills to system RAM (CPU/PCIe offload) and drops to the degraded figure.

Model sizeFP16Q8Q4
3B 122.6-146.2 t/s
≤222k ctx · then ~4.9-5.8 (slow)
245.3-292.3 t/s
≤268k ctx · then ~9.7-11.6 (slow)
438-522 t/s
≤288k ctx · then ~17.4-20.7 (slow)
7-8B 46-54.8 t/s
≤35k ctx · then ~1.8-2.2 (slow)
92-109.6 t/s
≤96k ctx · then ~3.7-4.4 (slow)
164.3-195.8 t/s
≤123k ctx · then ~6.5-7.8 (slow)
13B won't run 56.6-67.5 t/s
≤46k ctx · then ~2.2-2.7 (slow)
101.1-120.5 t/s
≤81k ctx · then ~4-4.8 (slow)
32B won't run won't run 41.1-48.9 t/s
≤10k ctx · then ~1.6-1.9 (slow)
70B won't run won't run won't run
120B won't run won't run won't run
400B won't run won't run won't run

Estimate: tok/s ≈ bandwidth ÷ (params × bytes/param) × efficiency (NVIDIA). Cyan = fits ~22.1 GB usable. Beyond "ctx", KV spills to system RAM (CPU/PCIe offload) (rough degraded speed shown).

What else you need for a full build

Minimum supporting components for a desktop built around the NVIDIA GeForce RTX 4090. Guidance — exact needs depend on your case, other parts and how much you offload to system RAM.

Power supply
≥ 850W · 80+ Gold
Headroom over the 450W card plus the rest of the build.
System RAM
≥ 32 GB · 64 GB recommended
At least the 24 GB of VRAM; 2× lets you CPU-offload models larger than VRAM.
Motherboard
1× free PCIe x16 slot (4.0/5.0)
A full-length x16 electrical slot; PCIe 4.0 is ample, 5.0 future-proofs.
CPU
Any modern multi-core
The GPU runs inference; more cores and memory channels only matter when offloading to RAM.
Storage
≥ 1 TB NVMe SSD
Weights are large — a 70B model at Q4 is ~40 GB; budget for a few.
PSU connector
12V-2×6 / 12VHPWR (or 3× 8-pin)
High-power NVIDIA cards use the 16-pin 12V-2×6 power connector.
Real-world measurements (2)
ModelQuantRuntimetok/sCtxSource
Gemma 4 26B Q4_K_M Ollama 149.6 4k blog ↗
Gemma 4 31B Q4_K_M Ollama 7.8 4k blog ↗

Run popular models on the NVIDIA GeForce RTX 4090

Videos about the NVIDIA GeForce RTX 4090

LLMs on RTX 4090 Laptop vs Desktop 🤯 not even close!

4090 Local AI Server Benchmarks

More hardware like the NVIDIA GeForce RTX 4090

About the NVIDIA GeForce RTX 4090

NVIDIA GeForce RTX 4090 is a discrete GPU from NVIDIA on Ada Lovelace, 2022. For local genAI the numbers that matter are its 24 GB (how big a model + context fits) and 1008 GB/s bandwidth (generation speed). Plan for a ~850W power supply.

Frequently asked questions

What size LLM can the NVIDIA GeForce RTX 4090 run?

With ~22.1 GB usable it runs up to ~32.5B at Q4 (~9.5B at FP16) at 8k context. Bigger spills into system RAM (CPU/PCIe offload) and slows sharply.

How many tokens/sec does the NVIDIA GeForce RTX 4090 generate?

Generation is bandwidth-bound. At 1008 GB/s an 8B model at Q4 runs ~164.3-195.8 tok/s (estimate).

Why does long context slow it down?

The KV cache for the context must also fit fast memory; once weights+KV exceed 22.1 GB it spills to system RAM (CPU/PCIe offload) and throughput collapses. Each row shows the max usable context.

What else do I need to run the NVIDIA GeForce RTX 4090?

Besides the card: a ≥850W 80+ Gold power supply, at least 24GB of system RAM (2× the 24GB VRAM is ideal for CPU offload), a free PCIe x16 slot, any modern multi-core CPU, and a ≥1TB NVMe SSD for model weights.