NVIDIA GeForce
NVIDIA GeForce RTX 5090 32GB
Architecture
Blackwell
VRAM
32 GB
Compare NVIDIA RTX GPUs by VRAM, see which local models fit, and check Ollama or LM Studio compatibility using deterministic AIPCFit data.
Use this hub when you are choosing a GPU for local LLMs. Use the PC checker when you want results for your exact GPU, RAM, operating system, and local AI workload.
NVIDIA RTX directory
NVIDIA GeForce
Architecture
Blackwell
VRAM
32 GB
NVIDIA GeForce
Architecture
Blackwell
VRAM
16 GB
NVIDIA GeForce
Architecture
Blackwell
VRAM
16 GB
NVIDIA GeForce
Architecture
Blackwell
VRAM
16 GB
NVIDIA GeForce
Architecture
Ada Lovelace
VRAM
24 GB
NVIDIA GeForce
Architecture
Ada Lovelace
VRAM
16 GB
NVIDIA GeForce
Architecture
Ada Lovelace
VRAM
16 GB
NVIDIA GeForce
Architecture
Ada Lovelace
VRAM
12 GB
NVIDIA GeForce
Architecture
Ada Lovelace
VRAM
16 GB
NVIDIA GeForce
Architecture
Ada Lovelace
VRAM
8 GB
NVIDIA GeForce
Architecture
Ampere
VRAM
24 GB
NVIDIA GeForce
Architecture
Ampere
VRAM
12 GB
Choosing hardware
Start with dedicated VRAM, then confirm model fit against the exact model size, quantization, context length, runtime overhead, and offloading behavior you plan to use. System RAM can still matter when layers or model data are handled outside GPU memory.
Higher VRAM generally gives more room for larger variants or longer contexts, but it does not guarantee that a model always fits or runs well. Treat these links as planning tiers, not official requirements.
8GB VRAM guide
Planning tier for smaller quantized local LLM variants and careful headroom.
12GB VRAM guide
Planning tier for common quantized local LLM work with more room for runtime overhead.
16GB VRAM guide
Planning tier for common quantized local LLM work with more room for runtime overhead.
24GB VRAM guide
Planning tier for larger variants or more comfortable local AI memory headroom.
32GB VRAM guide
Planning tier for larger variants or more comfortable local AI memory headroom.
48GB VRAM guide
Planning tier for larger variants or more comfortable local AI memory headroom.
VRAM planning
AIPCFit does not rank one GPU as universally best. Use these VRAM tiers as practical planning ranges, then check the specific model, quantization, context length, runtime overhead, system RAM, and software path before deciding.
Compare local LLMs by hardware fitBest treated as a careful planning tier for smaller quantized models, tighter context windows, and conservative runtime settings.
Adds more practical headroom for common small-to-mid local models, especially when quantization and context length are chosen carefully.
Gives stronger flexibility for mid-size models, larger quantizations, and more comfortable local development experiments.
Supports broader experimentation with larger models, more context, and extra runtime headroom, while still depending on the exact model and settings.
NVIDIA RTX catalog
6GB VRAM
2 catalog GPUs
8GB VRAM
15 catalog GPUs
10GB VRAM
1 catalog GPU
11GB VRAM
1 catalog GPU
12GB VRAM
8 catalog GPUs
16GB VRAM
7 catalog GPUs
24GB VRAM
3 catalog GPUs
32GB VRAM
1 catalog GPU
Model-fit discovery
These examples use dedicated VRAM only and do not infer system RAM. Open a GPU page or the compatibility hub for model-specific pages.
Representative GPU
Representative GPU
Representative GPU
Representative GPU
Representative GPU
How much VRAM do local LLMs need?
A larger model or less compressed quantization needs more memory. Runtime overhead, context length, KV cache, other GPU work, system RAM, and CPU offloading can change what is practical. File size alone is not the same as full runtime VRAM.
AIPCFit uses structured hardware data and deterministic compatibility rules.
Results are planning guidance, not performance guarantees.