If you want to run large language models locally without building a multi-GPU tower, two small computers dominate the conversation: NVIDIA's DGX Spark, built on its GB10 Grace Blackwell Superchip, and mini PCs using AMD's Ryzen AI Max+ 395 processor (codenamed Strix Halo). Both pair a powerful chip with up to 128GB of unified memory, which lets them load models far too big for a typical graphics card.
But they are very different machines. Forge PC has not tested either system. This comparison uses NVIDIA's and AMD's official specifications, early llama.cpp results reported by Hardware Corner, analysis from Notebookcheck, and UK prices from Scan and Framework, which are approximate and were checked on 14 November 2026.
Specifications compared
| NVIDIA DGX Spark | AMD Ryzen AI Max+ 395 |
|---|
CPU | 20-core Arm (Grace) | 16 Zen 5 cores, 32 threads, up to 5.1GHz |
Graphics | Blackwell GPU | Radeon 8060S, 40 graphics cores, up to 2900MHz |
NPU | None listed | Up to 50 TOPS |
AI performance (maker's figure) | Up to 1 petaFLOP at FP4 | Up to 126 TOPS overall; up to 60 TFLOPS FP16 |
Memory | 64GB (OEM partners only) or 128GB unified | Up to 128GB LPDDR5x-8000, 256-bit |
Memory bandwidth | 273GB/s (NVIDIA spec, via Laptop Deals UK) | About 256GB/s (calculated from 256-bit LPDDR5x-8000) |
Operating system | DGX OS (Ubuntu-based Linux) | Windows 11, Ubuntu or RHEL (x86-64) |
Clustering | Up to four units via ConnectX-7 | Not a built-in feature |
The AMD bandwidth figure is our own arithmetic: a 256-bit bus at 8,000MT/s works out at 256GB/s. AMD's product page lists the bus width and memory speed rather than a bandwidth number. The raw bandwidth figures are close, so the bigger differences are in compute and software.
Memory and model size
Both systems can hold very large models in memory. NVIDIA says the DGX Spark can run inference on models of up to 200 billion parameters and fine-tune models of up to 70 billion. Clustering two units with a direct cable doubles the memory pool to 256GB, according to NVIDIA, and four units need an external 200GbE switch.
On the AMD side, the amount of memory depends on the mini PC you buy, up to 128GB, and it is soldered, so it cannot be upgraded later. Framework, for example, says its Desktop's LPDDR5x is "non-upgradeable". Buy the 128GB version if you plan to run the largest open models.
How much memory you actually need depends on model size and quantisation. Our RAM for local AI guide explains the maths.
Hardware Corner published early llama.cpp results in October 2025, comparing the two chips on OpenAI's gpt-oss 120B model in MXFP4 format:
Test (gpt-oss 120B MXFP4) | DGX Spark | AMD Strix Halo |
|---|
Prompt processing (pp2048), tokens/s | 1,723.07 | 339.87 |
Token generation (tg32), tokens/s | 38.55 | 34.13 |
These figures are Hardware Corner's, not ours, and they are early results that may have changed with software updates. The pattern is still useful. The DGX Spark is several times faster at processing long prompts, which matters for big documents, retrieval-augmented generation and coding agents with large contexts. Token generation, the speed at which an answer appears, is only marginally faster than Strix Halo, because both are limited by similar memory bandwidth.
Software: CUDA vs x86 flexibility
This is the deciding factor for many buyers.
DGX Spark runs DGX OS, an Ubuntu-based Linux distribution, with NVIDIA's AI software stack preinstalled. You get CUDA, which Notebookcheck calls "almost unavoidable" if you are preparing code for large data centres. But it is an Arm system. Laptop Deals UK points out that your software and containers need to support Linux on Arm, and CUDA support alone is not enough.
Ryzen AI Max+ 395 is a standard x86 PC chip. AMD lists Windows 11, Ubuntu and RHEL support. Notebookcheck says it "slips seamlessly into the Windows ecosystem" and runs legacy applications, but that AMD's ROCm software does not yet quite match NVIDIA's stack for compatibility in many specialist applications. It can also be a normal desktop for everyday use and gaming.
If you want to run local AI on AMD hardware, our AMD local AI guide and our llama.cpp vs Ollama comparison cover the software.
UK prices
Prices were checked on 14 November 2026 and are approximate:
NVIDIA DGX Spark 128GB with 4TB SSD: about £6,399 at Scan.
Framework Desktop DIY Edition, Ryzen AI Max+ 395 with 128GB: £3,449 for the system on Framework's UK store, including VAT, but listed as out of stock. The 64GB version is £1,959, also out of stock. Storage and other options in the DIY Edition are configured separately.
Other mini PCs using the Ryzen AI Max+ 395 are available from various makers, and the same Grace Blackwell GB10 chip is sold in systems from NVIDIA's partners, such as the Gigabyte AI TOP ATOM in our database. NVIDIA says hardware support for a DGX Spark comes from the vendor that sold it. Our RTX Spark desktops story and Ryzen AI Max PRO 495 story cover newer systems on the way.
Which should you buy?
Choose the DGX Spark if:
You are developing for NVIDIA hardware and want CUDA and NVIDIA's AI stack out of the box.
Your work involves long prompts and large contexts, where its prompt-processing speed shines.
You may want to cluster two or four units, or fine-tune models.
Choose a Ryzen AI Max+ 395 mini PC if:
You mainly want to run (inference) large models and value cost: Notebookcheck calls the AMD chip "often more cost-effective" for pure inference.
You want Windows, x86 software compatibility, or a machine that doubles as an everyday desktop.
You are happy with generation speed similar to the Spark and slower prompt processing.
If neither fits, our best PC for local AI guide covers graphics card-based builds, and our Mac vs PC for local AI comparison covers Apple's alternatives.
Our verdict
The DGX Spark is the better developer machine for NVIDIA's ecosystem, with far faster prompt processing in Hardware Corner's early results, CUDA out of the box and the option to cluster. A Ryzen AI Max+ 395 mini PC costs roughly half as much in the UK, generates text at a similar speed, and is a normal x86 Windows or Linux PC, which makes it the better value for running models rather than building for NVIDIA hardware. Check stock before you plan around Framework's AMD system, which was out of stock on 14 November 2026. Forge PC has not tested either machine; specifications come from NVIDIA and AMD, and performance figures are Hardware Corner's.