Mac or PC for local LLMs: unified memory against a graphics card

A graphics card has fast memory and not much of it. A Mac has more memory than any consumer card and slower access to it. That one trade explains almost everything about which machine suits which model, and it is why the answer to "Mac or PC" is a question about the size of the model you want.

Where each one wins

Capacity goes to the Mac. Apple Silicon shares one pool of memory between the processor and the GPU, so a machine with 128 GB can load a model that would need several graphics cards. An M4 Max has 546 GB/s and an M3 Ultra has 819, and the table shows what that buys against two of the best consumer cards.

Device	Memory	Bandwidth	Qwen3 32B	Llama 3.3 70B Instruct	GPT OSS 120B
GeForce RTX 4090	24 GB	1008 GB/s	37.1 t/s	does not fit	does not fit
GeForce RTX 5090	32 GB	1792 GB/s	66.0 t/s	does not fit	does not fit
Apple M4 Max 128GB	128 GB	546 GB/s	18.6 t/s	8.9 t/s	68.4 t/s
Apple M3 Ultra 256GB	256 GB	819 GB/s	27.9 t/s	13.4 t/s	103 t/s
Ryzen AI Max+ 395 128GB	128 GB	256 GB/s	7.3 t/s	3.5 t/s	26.8 t/s

Speed goes to the card whenever the model fits. A 24 GB RTX 4090 reads memory about twice as fast as an M4 Max, so on a 32B model it answers at about twice the speed, and the gap is wider against an M4 Pro. The moment a model passes the card's memory the comparison flips, because the Mac loads it and the card does not. The 70B column above is the example: it fits a 128 GB Mac and no single consumer card.

Mixture-of-experts models narrow the gap. They read only a few experts per token, so a Mac with 128 GB of memory runs a 120B model at a speed that feels interactive, and the same model on a PC needs 64 GB of fast system RAM next to a mid-range card, as the MoE guide describes.

The catch is prompt processing

Everything above is generation speed, which is limited by memory bandwidth. Reading your prompt is limited by compute, and here Apple chips lag. For an 8B model our estimates put an M4 Max near 344 tokens per second of prompt processing against about 7,808 on an RTX 4090. For a chat with short messages you will not notice. For a 20,000-token document or a coding assistant that re-reads your repository on every turn, the Mac waits several seconds longer before it starts to answer, and that is the complaint Mac users raise most often.

Two practical things help. Keep the conversation going in one session so the runtime reuses its cached prompt, and use a smaller model for the long-context jobs. Neither erases the gap.

The practical differences

A Mac is quiet, small and draws a fraction of the power of a desktop with a 450 W card, which matters if it sits on your desk all day. It also works the first time, with Ollama, LM Studio and MLX all running natively. In return you cannot upgrade it, the memory is soldered at the price you paid on the day, and macOS caps how much of it the GPU may use. By default that is about two thirds to three quarters, and on recent versions you can raise the limit with sudo sysctl iogpu.wired_limit_mb=N, a setting that resets at reboot. We account for the default in the calculator and flag models that only fit with the limit raised.

A PC is faster per dollar, upgradeable, and the ecosystem is wider: every new runtime and every CUDA optimization lands there first. It is also louder, hotter and limited by VRAM, and a second card is the only way to grow. The multi-GPU guide covers what that costs you.

A third option sits between them. Mini PCs built on AMD's Ryzen AI Max+ 395 offer up to 128 GB of shared memory at 256 GB/s, and NVIDIA's DGX Spark has 128 GB at 273 GB/s. They load the same large models as a Mac for less money and with less software polish.

Machines by budget

These are whole computers, not cards, ranked by the strongest model each one runs well. Prices are those of the cheapest configuration with that amount of memory.

Under $1,500

Device	Memory	Best all-round model	Speed
Apple M1 Max 32GB $1,400, whole machine, 400 GB/s	32 GB	Gemma 4 12B Instruct at Q8_0	20.7 t/s
Ryzen AI Max+ 395 32GB $1,299, whole machine, 256 GB/s	32 GB	GPT OSS 20B at MXFP4	35.3 t/s
Apple M4 Pro 24GB $1,399, whole machine, 273 GB/s	24 GB	GPT OSS 20B at MXFP4	45.1 t/s
Apple M4 24GB $799, whole machine, 120 GB/s	24 GB	GPT OSS 20B at MXFP4	19.8 t/s
Apple M4 32GB $999, whole machine, 120 GB/s	32 GB	GPT OSS 20B at MXFP4	19.8 t/s

$1,500 to $3,000

Device	Memory	Best all-round model	Speed
Ryzen AI Max+ 395 96GB $1,999, whole machine, 256 GB/s	96 GB	GPT OSS 120B at MXFP4	26.8 t/s
Ryzen AI Max+ 395 128GB $2,199, whole machine, 256 GB/s	128 GB	GPT OSS 120B at MXFP4	26.8 t/s
Apple M1 Ultra 64GB $2,400, whole machine, 800 GB/s	64 GB	Qwen3.5 27B at Q8_0	20.2 t/s
Ryzen AI Max+ 395 64GB $1,699, whole machine, 256 GB/s	64 GB	Qwen3.5 35B A3B at Q8_0	27.7 t/s
Apple M1 Max 64GB $1,900, whole machine, 400 GB/s	64 GB	Qwen3.5 35B A3B at Q8_0	51.8 t/s

$3,000 to $5,000

Device	Memory	Best all-round model	Speed
Apple M1 Ultra 128GB $3,400, whole machine, 800 GB/s	128 GB	GPT OSS 120B at MXFP4	100.1 t/s
Apple M2 Max 96GB $3,699, whole machine, 400 GB/s	96 GB	GPT OSS 120B at MXFP4	50 t/s
Apple M3 Ultra 96GB $3,999, whole machine, 819 GB/s	96 GB	GPT OSS 120B at MXFP4	102.5 t/s
NVIDIA DGX Spark 128GB $3,999, whole machine, 273 GB/s	128 GB	GPT OSS 120B at MXFP4	37 t/s
Apple M3 Max 96GB $4,199, whole machine, 400 GB/s	96 GB	GPT OSS 120B at MXFP4	50 t/s

$5,000 to $10,000

Device	Memory	Best all-round model	Speed
Apple M3 Ultra 512GB $9,499, whole machine, 819 GB/s	512 GB	MiniMax M2.7 at Q8_0	25.3 t/s
Apple M3 Ultra 256GB $5,599, whole machine, 819 GB/s	256 GB	Qwen3.5 122B A10B at Q8_0	38.6 t/s
Apple M2 Ultra 192GB $6,599, whole machine, 800 GB/s	192 GB	Qwen3.5 122B A10B at Q8_0	37.7 t/s

The pattern is consistent. Each step up in price mostly buys memory, and memory is what opens the way to the next model size, so the strongest pick changes in jumps and not smoothly. The second thing to read in the table is the speed column: a machine that loads a large model and answers at 10 tokens per second is a different purchase from one that answers at 60, and the price does not tell you which you are getting.

How to choose

If the model you want fits in 24 GB, buy the card: it is faster and cheaper. If you want 70B or larger and you chat more than you feed it long documents, buy the memory. If you want both, a PC with a card for everyday work is worth more than a Mac that does the big jobs slowly. Check the exact pairing in the calculator for the models you actually plan to use before you spend anything.