NV NVIDIA Consumer Discrete GPU

NVIDIA GeForce RTX 5090

Runs downloadable LLMs up to ~45B at Q4, ~292-348 tok/s on an 8B model.

NVIDIA GeForce RTX 5090
32GB
VRAM
1792
GB/s bandwidth
45B
Max model (Q4)
$1,999
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 218-259.8 t/s
≤335k ctx · then ~4.9-5.8 (slow)
436.1-519.7 t/s
≤381k ctx · then ~9.7-11.6 (slow)
778.7-928 t/s
≤401k ctx · then ~17.4-20.7 (slow)
7-8B 81.8-97.4 t/s
≤91k ctx · then ~1.8-2.2 (slow)
163.5-194.9 t/s
≤152k ctx · then ~3.7-4.4 (slow)
292-348 t/s
≤179k ctx · then ~6.5-7.8 (slow)
13B 50.3-60 t/s
≤12k ctx · then ~1.1-1.3 (slow)
100.6-119.9 t/s
≤91k ctx · then ~2.2-2.7 (slow)
179.7-214.2 t/s
≤126k ctx · then ~4-4.8 (slow)
32B won't run won't run 73-87 t/s
≤38k 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 ~29.4 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 5090. Guidance — exact needs depend on your case, other parts and how much you offload to system RAM.

Power supply
≥ 1000W · 80+ Gold
Headroom over the 575W card plus the rest of the build.
System RAM
≥ 32 GB · 64 GB recommended
At least the 32 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.

Run popular models on the NVIDIA GeForce RTX 5090

Videos about the NVIDIA GeForce RTX 5090

Not even close‼️LLMs on RTX5090 vs others

5090 Local AI GPU Performance Meta-Analysis

More hardware like the NVIDIA GeForce RTX 5090

About the NVIDIA GeForce RTX 5090

NVIDIA GeForce RTX 5090 is a discrete GPU from NVIDIA on Blackwell, 2025. For local genAI the numbers that matter are its 32 GB (how big a model + context fits) and 1792 GB/s bandwidth (generation speed). Plan for a ~1000W power supply.

Frequently asked questions

What size LLM can the NVIDIA GeForce RTX 5090 run?

With ~29.4 GB usable it runs up to ~45B at Q4 (~13B 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 5090 generate?

Generation is bandwidth-bound. At 1792 GB/s an 8B model at Q4 runs ~292-348 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 29.4 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 5090?

Besides the card: a ≥1000W 80+ Gold power supply, at least 32GB of system RAM (2× the 32GB VRAM is ideal for CPU offload), a free PCIe x16 slot, any modern multi-core CPU, and a ≥1TB NVMe SSD for model weights.