Is NVIDIA, AMD or a Mac better for local LLMs?

It depends on the model size you want. For models up to about 32B that fit in one card, a desktop GPU gives the most speed per dollar because it has far more memory bandwidth. To run Qwen3.8-27B at Q4, the cheapest NVIDIA card we track is the GeForce RTX 3090 ($1,499) and the cheapest AMD card is the Radeon RX 7900 XTX ($999). For 70B-class models and big MoE models, one unified-memory machine (a 128 GB Mac or mini PC) is the simplest option, since no single consumer GPU we track has enough VRAM.

Is AMD good for running local LLMs?

AMD Radeon cards work in llama.cpp-based apps (such as LM Studio and Ollama) through ROCm or Vulkan, and they are often cheaper per GB of VRAM. The cheapest 24 GB AMD card we track is the Radeon RX 7900 XTX at $999, versus the GeForce RTX 3090 at $1,499 for NVIDIA. NVIDIA's CUDA still has the widest app support, and some tools add AMD support later or not at all.

Is a Mac good for local LLMs?

Yes for big models and quiet, simple setups: Macs share one pool of memory between CPU and GPU, and we count about 75% of it as usable. The cheapest current Mac we track that fits Llama 3.3 70B at Q4 is the MacBook Pro M5 Max 128GB ($6,699). Bandwidth per dollar is lower than a desktop GPU, so models that fit on a GPU usually run faster there.

What is the cheapest way to get 24 GB for local AI?

NVIDIA GeForce: the GeForce RTX 3090 at $1,499; AMD Radeon: the Radeon RX 7900 XTX at $999; Intel Arc: the Intel Arc Pro B70 32GB at $949; Apple Mac: the Mac mini M6 32GB at $1,799; Unified-memory mini PCs: the Framework Desktop (Ryzen AI Max+ 395) 128GB at $3,449. Prices are dated list-price snapshots from our data.

Which platform gives the most memory per dollar?

Among current devices we track, the Radeon RX 7600 XT at about $21 per usable GB ($329 for 16 GB).

Can I run a 70B model on NVIDIA or AMD?

Not on one consumer card we track: Llama 3.3 70B needs about 50.2 GB at Q4, more than the 32 GB maximum. Two 24 GB cards (NVIDIA or AMD) or a 128 GB unified-memory machine will hold it. See our ready-made builds.

NVIDIA vs AMD vs Mac for local LLMs

Which platform to buy for running open models at home: NVIDIA GeForce, AMD Radeon and Intel Arc graphics cards against Apple Macs and unified-memory mini PCs, compared on usable memory, price per GB, bandwidth and what each can run. Every number comes from the 82 devices we track.

NVIDIA GeForce: 34 tracked
AMD Radeon: 17 tracked
Intel Arc: 5 tracked
Apple Mac: 10 tracked
Unified-memory mini PCs: 3 tracked

Short answer: for models up to ~32B, a desktop GPU is fastest per dollar (cheapest that fits Qwen3.8-27B at Q4: GeForce RTX 3090, $1,499; AMD: Radeon RX 7900 XTX, $999). For 70B and larger, a unified-memory machine is the simplest single box (cheapest that fits Llama 3.3 70B: Framework Desktop (Ryzen AI Max+ 395) 128GB, $3,449).

The platforms at a glance

Price range covers current models (discontinued or reseller-priced devices are left out of ranges and picks). Graphics cards count all VRAM as usable; unified-memory machines about 75% (how we calculate).

Each platform's price range, most memory, most bandwidth, best value and biggest model at Q4
Platform	Price range	Most memory	Most bandwidth	Best $ per usable GB	Biggest model at Q4
NVIDIA GeForce34 tracked · dedicated VRAM, 100% usable	$249–$1,999current models	GeForce RTX 509032 GB usable	GeForce RTX 50901792 GB/s	GeForce RTX 5060 Ti 16GB$27 per usable GB	IQuest-Coder-V1-40B-Loop-Instruct~28.7 GB at Q4, on the GeForce RTX 5090
AMD Radeon17 tracked · dedicated VRAM, 100% usable	$269–$999current models	Radeon RX 7900 XTX24 GB usable	Radeon RX 7900 XTX960 GB/s	Radeon RX 7600 XT$21 per usable GB	ERNIE-4.5-VL-28B-A3B-PT~21.5 GB at Q4, on the Radeon RX 7900 XTX
Intel Arc5 tracked · dedicated VRAM, 100% usable	$219–$949current models	Intel Arc Pro B70 32GB32 GB usable	Intel Arc Pro B70 32GB608 GB/s	Intel Arc B580$21 per usable GB	IQuest-Coder-V1-40B-Loop-Instruct~28.7 GB at Q4, on the Intel Arc Pro B70 32GB
Apple Mac10 tracked · unified memory, ~75% usable	$899–$22,299current models	Mac Studio M5 Ultra 512GB384 GB usable	Mac Studio M5 Ultra 256GB1200 GB/s	Mac Studio M5 Ultra 256GB$49 per usable GB	Qwen3-Coder-480B-A35B-Instruct~335.8 GB at Q4, on the Mac Studio M5 Ultra 512GB
Unified-memory mini PCs3 tracked · unified memory, ~75% usable	$3,449–$6,950current models	Framework Desktop (Ryzen AI Max+ 395) 128GB96 GB usable	NVIDIA DGX Spark 64GB273 GB/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$36 per usable GB	Mistral-Small-4-119B-2603~84.2 GB at Q4, on the Framework Desktop (Ryzen AI Max+ 395) 128GB

Cheapest way to each memory size

The cheapest current device on each platform with at least that much usable memory. When no single card is big enough, the GPU columns show two cards (GPU cost only; you also need a board, PSU and case that take two cards, see the planner). Two cards add memory, not speed.

Cheapest device per platform for each usable memory size
Usable memory	NVIDIA GeForce	AMD Radeon	Intel Arc	Apple Mac	Unified-memory mini PCs
12 GB usable	GeForce RTX 3060 12GB$329 · 12 GB · 360 GB/s	Radeon RX 7600 XT$329 · 16 GB · 288 GB/s	Intel Arc B580$249 · 12 GB · 456 GB/s	Mac mini M6 16GB$899 · 12 GB · 153 GB/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 96 GB · 256 GB/s
16 GB usable	GeForce RTX 5060 Ti 16GB$429 · 16 GB · 448 GB/s	Radeon RX 7600 XT$329 · 16 GB · 288 GB/s	Intel Arc Pro B70 32GB$949 · 32 GB · 608 GB/s	Mac mini M6 24GB$1,299 · 18 GB · 170 GB/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 96 GB · 256 GB/s
24 GB usable	GeForce RTX 3090$1,499 · 24 GB · 936 GB/s	Radeon RX 7900 XTX$999 · 24 GB · 960 GB/s	Intel Arc Pro B70 32GB$949 · 32 GB · 608 GB/s	Mac mini M6 32GB$1,799 · 24 GB · 170 GB/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 96 GB · 256 GB/s
32 GB usable	GeForce RTX 5090$1,999 · 32 GB · 1792 GB/s	2× Radeon RX 7600 XT$658 for the two cards only · 288 GB/s each	Intel Arc Pro B70 32GB$949 · 32 GB · 608 GB/s	Mac mini M5 Pro 64GB$3,199 · 48 GB · 307 GB/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 96 GB · 256 GB/s
48 GB usable	2× GeForce RTX 3090$2,998 for the two cards only · 936 GB/s each	2× Radeon RX 7900 XTX$1,998 for the two cards only · 960 GB/s each	2× Intel Arc Pro B70 32GB$1,898 for the two cards only · 608 GB/s each	Mac mini M5 Pro 64GB$3,199 · 48 GB · 307 GB/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 96 GB · 256 GB/s
96 GB usable	—	—	—	MacBook Pro M5 Max 128GB$6,699 · 96 GB · 614 GB/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 96 GB · 256 GB/s

Cheapest device per platform for popular models

The cheapest single device on each platform that fits the model at Q4_K_M with room for ~4k context, with estimated decode speed (how we estimate speed; MoE models show an upper bound only).

Cheapest device per platform that runs each model at Q4, with estimated tokens per second
Model	NVIDIA GeForce	AMD Radeon	Intel Arc	Apple Mac	Unified-memory mini PCs
Qwen3-8B~6.7 GB at Q4	GeForce RTX 3050 8GB$249 · 27–38 tok/s	Radeon RX 7600$269 · 35–49 tok/s	Intel Arc B570$219 · 46–65 tok/s	Mac mini M6 16GB$899 · 14–25 tok/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 23–41 tok/s
gpt-oss-20b~15.6 GB at Q4	GeForce RTX 3090$1,499 · up to 365 tok/s	Radeon RX 7900 XT$899 · up to 312 tok/s	Intel Arc Pro B70 32GB$949 · up to 237 tok/s	Mac mini M6 24GB$1,299 · up to 62 tok/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · up to 94 tok/s
Qwen3-14B~11.3 GB at Q4	GeForce RTX 5060 Ti 16GB$429 · 30–43 tok/s	Radeon RX 7600 XT$329 · 19–27 tok/s	Intel Arc Pro B70 32GB$949 · 41–58 tok/s	Mac mini M6 24GB$1,299 · 9–15 tok/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 13–23 tok/s
Qwen3.8-27B~20.4 GB at Q4	GeForce RTX 3090$1,499 · 33–47 tok/s	Radeon RX 7900 XTX$999 · 34–48 tok/s	Intel Arc Pro B70 32GB$949 · 22–31 tok/s	Mac mini M6 32GB$1,799 · 5–8 tok/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 7–12 tok/s
Qwen3-32B~23.8 GB at Q4	GeForce RTX 5090$1,999 · 54–77 tok/s	none fits at Q4	Intel Arc Pro B70 32GB$949 · 18–26 tok/s	Mac mini M5 Pro 64GB$3,199 · 7–12 tok/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 6–10 tok/s
Llama-3.3-70B-Instruct~50.2 GB at Q4	none fits at Q4	none fits at Q4	none fits at Q4	MacBook Pro M5 Max 128GB$6,699 · 6–11 tok/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · 3–5 tok/s
gpt-oss-120b~82.5 GB at Q4	none fits at Q4	none fits at Q4	none fits at Q4	MacBook Pro M5 Max 128GB$6,699 · up to 159 tok/s	Framework Desktop (Ryzen AI Max+ 395) 128GB$3,449 · up to 66 tok/s

How to choose

Speed on models that fit: decode speed tracks memory bandwidth. Desktop GPUs have the most per dollar (up to 1792 GB/s on NVIDIA, 960 GB/s on AMD, 1200 GB/s on the fastest Mac we track).
Model size: memory decides what fits. No consumer card we track has more than 96 GB; unified-memory machines reach 384 GB usable.
Software: NVIDIA (CUDA) has the widest support across apps. AMD and Intel cards work in llama.cpp-based apps (ROCm, Vulkan, SYCL), but some tools support them later. Macs run llama.cpp (Metal) and MLX.
Power and setup: a GPU needs a PC around it (see ready-made builds); a Mac or mini PC is a complete system.

Go deeper: best GPU for local LLMs · best Mac for local LLMs · GPU vs GPU · by budget · VRAM calculator.

Quick answers

Is NVIDIA, AMD or a Mac better for local LLMs?: It depends on the model size you want. For models up to about 32B that fit in one card, a desktop GPU gives the most speed per dollar because it has far more memory bandwidth. To run Qwen3.8-27B at Q4, the cheapest NVIDIA card we track is the GeForce RTX 3090 ($1,499) and the cheapest AMD card is the Radeon RX 7900 XTX ($999). For 70B-class models and big MoE models, one unified-memory machine (a 128 GB Mac or mini PC) is the simplest option, since no single consumer GPU we track has enough VRAM.
Is AMD good for running local LLMs?: AMD Radeon cards work in llama.cpp-based apps (such as LM Studio and Ollama) through ROCm or Vulkan, and they are often cheaper per GB of VRAM. The cheapest 24 GB AMD card we track is the Radeon RX 7900 XTX at $999, versus the GeForce RTX 3090 at $1,499 for NVIDIA. NVIDIA's CUDA still has the widest app support, and some tools add AMD support later or not at all.
Is a Mac good for local LLMs?: Yes for big models and quiet, simple setups: Macs share one pool of memory between CPU and GPU, and we count about 75% of it as usable. The cheapest current Mac we track that fits Llama 3.3 70B at Q4 is the MacBook Pro M5 Max 128GB ($6,699). Bandwidth per dollar is lower than a desktop GPU, so models that fit on a GPU usually run faster there.
What is the cheapest way to get 24 GB for local AI?: NVIDIA GeForce: the GeForce RTX 3090 at $1,499; AMD Radeon: the Radeon RX 7900 XTX at $999; Intel Arc: the Intel Arc Pro B70 32GB at $949; Apple Mac: the Mac mini M6 32GB at $1,799; Unified-memory mini PCs: the Framework Desktop (Ryzen AI Max+ 395) 128GB at $3,449. Prices are dated list-price snapshots from our data.
Which platform gives the most memory per dollar?: Among current devices we track, the Radeon RX 7600 XT at about $21 per usable GB ($329 for 16 GB).
Can I run a 70B model on NVIDIA or AMD?: Not on one consumer card we track: Llama 3.3 70B needs about 50.2 GB at Q4, more than the 32 GB maximum. Two 24 GB cards (NVIDIA or AMD) or a 128 GB unified-memory machine will hold it. See our ready-made builds.

Buying guides · Best AMD GPU · Compare builds · How much VRAM do you need? · Tokens per second