AMD vs Intel vs Apple vs NVIDIA for local LLMs: which chip wins?

Four chip families dominate the conversation about running AI on your own hardware: AMD Ryzen AI, Intel Core Ultra, Apple M-series, and NVIDIA. Every buying guide eventually collapses into a brand argument — but the brands are a proxy for two decisions that actually decide whether a local large language model runs well: how the chip holds memory, and which software stack it speaks. Get those two right and the badge on the box matters far less than the spec sheet suggests.

This is a platform comparison, not a box-by-box shootout. If you want specific machines with live prices, jump to the hardware page or the mini PC buying guide. If you want to understand why one architecture pulls ahead for your workload, read on.

The one number that decides everything: memory

Before brand or benchmark, the model has to fit. A language model lives in fast memory while it runs, and the amount your chip can address caps the model you can host.

• A 7–14B model at 4-bit needs roughly 8–16 GB — any 32 GB box handles it.
• A 70B model at 4-bit needs around 40 GB — which is why every 70B-capable box ships 128 GB of unified memory, with up to 96 GB handed to the GPU.

That single requirement splits the four platforms into two memory models, and almost everything else follows from which one a chip uses. We size memory to models in more detail in what hardware you need to run an LLM locally.

Memory model one: unified memory (AMD, Apple, DGX Spark)

In a unified-memory design, the CPU and GPU share one big pool of RAM. There is no separate "video memory" to fill — the GPU can address most of the system memory directly, so a 128 GB box can load a 70B model that no consumer graphics card could hold.

This is the architecture behind AMD's "Strix Halo" (Ryzen AI Max+ 395) boxes — the Minisforum MS-S1 MAX, Beelink GTR9 Pro and Framework Desktop — as well as Apple's M-series Macs and NVIDIA's DGX Spark. They all trade peak bandwidth for capacity: you get an enormous, affordable pool, at "only" a few hundred GB per second.

• AMD Strix Halo boxes run ~256 GB/s of bandwidth on 128 GB of unified memory. That is the value sweet spot for a 70B in one affordable machine.
• NVIDIA DGX Spark pairs 128 GB unified (273 GB/s) with full CUDA — capacity and the best software stack in one box, which is rare.
• Apple Mac Studio M3 Ultra is the outlier: up to 256 GB unified at 819 GB/s — the highest memory bandwidth of any box we list. Unified capacity and near-discrete speed, which is why it is the fastest pure-inference machine here.

The catch for unified memory is bandwidth: it is generous but not GPU-class, so for a model that does fit a graphics card, a discrete GPU will still decode faster.

Memory model two: discrete VRAM (NVIDIA RTX)

A discrete GPU has its own memory soldered next to the chip, and it is blazing fast — but small. An RTX 5090 SFF has 32 GB of GDDR7 running at 1,792 GB/s, roughly seven times the bandwidth of a Strix Halo box. For any model that fits in 32 GB, nothing on this page decodes tokens faster.

The hard limit is capacity. A 70B model at 4-bit needs about 40 GB; it simply does not fit in 32 GB of VRAM. So the discrete-GPU path forks:

• Consumer/prosumer — the RTX 5090 SFF is the fastest box for ~30B-class models and anything under 32 GB. Past that ceiling you are stuck.
• Workstation/data-centre — the RTX PRO 6000 (96 GB GDDR7 ECC) and dual-card or DGX Station GB300 configurations restore the capacity, at quote-tier prices, and add the concurrency to serve many users at once.

If you are weighing a single fast GPU against a big unified box for a 70B, we put them head to head in can a mini PC run a 70B model?.

The software stack: CUDA, Metal, ROCm or oneAPI

Memory decides what you can run; the software stack decides how painlessly. This is where NVIDIA's lead is real and where the others differentiate.

• NVIDIA — CUDA. The broadest, most mature ecosystem by a wide margin. vLLM, TensorRT-LLM, and nearly every new model and serving framework target CUDA first. Crucially, it gives you the closest parity with the cloud: what you prototype on a managed GPU you can redeploy on a local NVIDIA box with little friction.
• Apple — Metal. MLX and llama.cpp are well-optimised for Apple Silicon, and the M3 Ultra's 819 GB/s bandwidth makes decode genuinely fast. The trade-off: it is Metal, not CUDA, so some CUDA-only tooling has no direct equivalent.
• AMD — ROCm / Vulkan. Steadily improving and the best value per gigabyte of capacity. llama.cpp via Vulkan and a maturing ROCm stack run 70B-class models well on Strix Halo. You sometimes do a little more setup than on CUDA, but the price-to-capacity ratio is unmatched.
• Intel — Arc / oneAPI. The smallest LLM ecosystem of the four, but Intel's real pitch is not raw speed — it is fleet manageability. vPro and the management features around the ASUS NUC 14 Pro AI make it the natural pick for IT shops standardising a managed estate.

The NPU myth: why "50 TOPS" rarely matters

Every Copilot+ laptop and most of these mini PCs advertise an NPU rated at "50 TOPS" or similar. It is the headline number on the box. For local LLM serving, it is largely irrelevant.

Local LLM inference runs on the integrated GPU (or a discrete GPU), not the NPU. The NPU is a low-power accelerator designed for small, bursty on-device tasks — background blur, live captions, the OS's own assistant features — not for streaming tokens from a 14B chat model under load. The serving stacks that matter (llama.cpp, Ollama, vLLM, MLX) target the GPU.

When you compare two boxes for local AI, ignore the TOPS figure. Look at memory capacity, memory bandwidth, and which software stack the GPU speaks. Those three numbers predict real-world tokens per second; the NPU rating predicts almost nothing about it.

This is why a chip with a modest NPU but generous unified memory will out-serve a thinner box with a bigger TOPS sticker every time.

So which chip should you pick?

There is no single winner — the right platform follows your use-case.

• Best value for one 70B, one box → AMD Strix Halo. The Minisforum and Beelink boxes give you 128 GB of unified memory and a real 70B at the lowest price. ROCm asks a little patience; the savings are large.
• Fastest inference, quietest box → Apple M-series. The Mac Studio M3 Ultra's 819 GB/s bandwidth makes it the speed king for pure inference, runs near-silent, and scales to 256 GB. Metal, not CUDA.
• Cloud parity and the broadest tooling → NVIDIA. DGX Spark for unified capacity with full CUDA; RTX 5090 for raw speed under 32 GB; RTX PRO 6000 and DGX Station for 70B-plus and many concurrent tenants.
• A managed fleet across your whole organisation → Intel Core Ultra. The ASUS NUC 14 Pro AI and vPro give IT the manageability to standardise dozens of identical, governed boxes.

Most small businesses do not need to agonise over this. A 7–14B model fits any 32 GB box from any vendor, so the entry tier is genuinely vendor-agnostic — pick on warranty, support and price, which the hardware page tracks live.

A note on prices and posture

AI hardware prices have been volatile through 2026 thanks to the DRAM and HBM shortage, so treat every figure as indicative and confirm the live price before ordering — the hardware page is re-verified monthly and carries the disclosed affiliate links. As an indicative range, the entry tier starts around ~€490 and a performance box runs from ~€2,400+, both flagged as approximate.

And whichever chip you choose, remember it is only half the product. A raw model on a fast box is a confidential database that will read its own secrets aloud to anyone who asks. What makes a local AI safe to put in front of customers is the governance layer — the boundary between your internal knowledge and the public agent — and that is hardware-independent. It is the company's bet that sovereignty starts with where the model runs, but the governance is what closes the loop. We make the case in the product is governance and explain the model in sovereign AI, explained.

So pick the platform that fits your model, then put a governed core on it. To compare the actual machines, see the mini PC buying guide or the full hardware shortlist.