Local AI buying guide
DGX Spark vs Strix Halo vs Mac Studio: a UK local-AI comparison
Three 128GB desktop routes, different software requirements. Compare exact configurations, usable memory and UK buying costs before choosing.
By Laptop Deals UK · Updated
The short answer: choose the software path first
Choose a DGX Spark candidate when your project needs an NVIDIA development environment and supports Linux on Arm. Consider a Ryzen AI Max+ desktop when you need Windows or an x86 workflow with a supported AMD runtime. Choose Mac Studio when your tools and models suit Apple Silicon, MLX or Metal. None is a universal winner simply because it has 128GB memory.
This comparison uses published specifications and dated UK retailer observations, not hands-on tests. We have not established matched inference speeds for these configurations.
Exact configurations, not family starting models
| Configuration | Relevant software route | Main buying constraint |
|---|---|---|
| NVIDIA DGX Spark Founder's Edition 128GB / 4TB self-encrypting SSD / 240W supply | NVIDIA GB10, Arm CPU and NVIDIA GPU; NVIDIA DGX OS. Confirm Linux-on-Arm packages and containers as well as CUDA compatibility. | A desktop, not a conventional x86 Windows PC. The selected UK retailer quote below includes no optional peripherals. |
| HP Z2 Mini G1a 8T219EA#ABU · Ryzen AI Max+ PRO 395 / 128GB / 2TB / Windows 11 Pro | x86 Windows workstation with Radeon 8060S integrated graphics. Use a supported AMD/HIP or Vulkan implementation for your exact model and driver. | Onboard memory cannot be upgraded. The configuration is useful to compare, but it is not an available offer at our latest stock check. One-year on-site limited warranty. |
| Mac Studio M5 Max 18-core CPU / 40-core GPU / 128GB / 2TB | macOS, with MLX or llama.cpp/Metal for supported models. NVIDIA CUDA packages need another implementation. | Select this exact memory, chip and storage in the configurator. Memory and internal storage are chosen at purchase; budget for peripherals. See Apple for a final configured total and delivery date. |
Desktop configuration details and their specification dates · GB10 retailer comparison
A UK purchase comparison needs more than the headline price
DGX Spark at CyberPowerPC UK
£6,099.00
Includes VAT and UK-mainland delivery.
In stock at check · CyberPowerPC UK
Price checked .
Before you buy: 128GB unified memory / 4TB self-encrypting SSD, 240W supply and one-year NVIDIA warranty, with no optional accessories selected. Includes UK-mainland delivery; dispatch was advertised for 11 November, with same-business-day dispatch for orders confirmed by 2pm. Confirm the dispatch date and whether a suitable UK mains lead is included: the selected Cable field says None. Display and peripherals cost extra.
Before you buy: Check the exact configuration, current purchase terms and delivery date with the retailer before ordering.
View at CyberPowerPC UKThe quote is for the selected 128GB/4TB machine, not an unrelated starting model. It does not establish that DGX Spark is the best-value computer for every workload. For other shops and any newer observations, use the UK GB10 price comparison.
We do not subtract cashback, trade-in value or a hypothetical business VAT recovery from a consumer budget. For the HP and Apple configurations, compare a genuinely orderable exact quote rather than treating an unavailable discount or a base Mac Studio price as equivalent. Storage also differs: the Spark here has 4TB, while these HP and Apple examples have 2TB.
128GB is not 128GB of free model space
All three routes share memory between system and GPU work. The OS, applications, model weights, runtime buffers and context/KV-cache compete for that pool. AMD documents up to 96GB graphics allocation for 128GB Ryzen AI Max systems, subject to the exact firmware and runtime; do not copy that limit onto GB10 or Apple Silicon. Equally, do not assume either of those makes its whole advertised pool available to weights.
NVIDIA specifies 273GB/s memory bandwidth for DGX Spark. Bandwidth, peak FP4 compute and NPU TOPS are not interchangeable benchmarks. Long prompt processing, token generation, image generation and training can favour different hardware or software. Work through the 128GB memory budget before paying for capacity you may not use.
What would justify choosing one over another?
- CUDA-based development: check whether the required package, extension and container support GB10 and Arm. An x86-only dependency can prevent an otherwise suitable NVIDIA workflow.
- Windows workstation work alongside local models: the HP combines workstation features with shared memory. Confirm your AI runtime separately; workstation certification is not an LLM performance result.
- A macOS development desk: the Mac Studio can suit an existing MLX/Metal workflow. Check model conversions, unsupported operators and runtime allocation before switching from another platform.
- Private chat or coding with smaller models: first try a compatible machine you already own. A supported lower-memory computer can be better value than buying a large pool for a workload that does not need it.
Ask for measurements naming the exact machine, power mode, OS, runtime version, model, quantisation, context and batch size. Compare prompt processing and generation separately. A vendor's multi-machine demonstration is not a single-desktop speed promise.
Sources and next comparisons
- NVIDIA DGX Spark hardware documentation
- AMD variable graphics memory guidance
- MLX documentation and llama.cpp backends
What fits in 128GB? · Observed non-Amazon price history · Portable 128GB alternatives