DGX Spark vs Strix Halo: 128 GB Local AI Desktops Compared
DGX Spark vs Strix Halo: both are 128 GB unified-memory desk boxes. Spark (NVIDIA GB10, CUDA, Arm, DGX OS, 4 TB) is the faster CUDA/vLLM machine; D-Central lists it at $9,449 CAD configured. Strix Halo (Ryzen AI Max+ 395, x86, Windows or Linux, ROCm/Vulkan) is the cheaper, more familiar PC; we list a sourcable 128 GB box at $7,349 CAD. We do not sell AMD’s Micro Center-exclusive “Ryzen AI Halo.” A custom GPU tower is the third path when you want discrete GDDR and a card you can swap.
This is a Canadian merchant comparison, dated 24 August 2026. Specs are from NVIDIA and AMD product documentation. Throughput claims below are attributed to named third-party reviews, not our lab. Credit NVIDIA, AMD, Ollama, llama.cpp, and vLLM.
Same job, different computers
NVIDIA created the category: a ~1-liter box with enough coherent memory to hold large open-weight models. AMD’s answer is Strix Halo — the same 128 GB class on x86. StorageReview’s July 2026 Halo review and Tom’s Hardware’s June 2026 coverage put the official Halo at US$3,999 with 2 TB, against Spark Founders around US$4,699 with 4 TB. Canadian retail Spark (bare) on this date sat roughly $7,650–$8,450. Our lists include first-boot setup in Laval, Quebec.
| NVIDIA DGX Spark | AMD Strix Halo 128 GB (sourcable) | |
|---|---|---|
| SoC | GB10 Grace Blackwell | Ryzen AI Max+ 395 (Strix Halo) |
| Memory | 128 GB LPDDR5X unified, 273 GB/s | 128 GB LPDDR5X unified (soldered) |
| Storage (typical) | 4 TB NVMe | 2 TB class (confirm on the order) |
| CPU ISA | Arm (20-core Grace) | x86-64 (16C/32T Zen 5) |
| GPU software | CUDA, TensorRT-LLM, DGX OS | ROCm, Vulkan, Windows or Linux |
| Power (rated) | ~240 W system / 140 W GB10 TDP | ~120 W sustained APU (chassis-dependent) |
| OS | DGX OS (Ubuntu-based) | Windows 11 or Linux |
| D-Central list | $9,449 CAD | $7,349 CAD |
What the reviews actually say
StorageReview (6 July 2026) found Spark held a roughly 2×–4× vLLM throughput lead in most of their sweeps, stretching further on prefill-heavy gpt-oss-120b, and narrowing only on some decode-heavy runs. The Register’s late-2025 Strix Halo vs Spark piece made the same directional point: Spark is the CUDA appliance; Strix Halo is the cheaper x86 box with the same memory ceiling. Treat those as dated lab snapshots, not a D-Central benchmark.
llama.cpp on Strix Halo often runs on Vulkan; ROCm is the AMD-native path and is not a drop-in CUDA. If your stack is vLLM + NVIDIA containers, Spark (or a discrete NVIDIA GPU) is the less painful machine.
When Spark is the right buy
- You need CUDA, NIM, TensorRT-LLM, or a DGX-shaped software path.
- You want 4 TB local weights and a 240 W plug.
- You can live on Arm64 Linux. Python/CUDA/Ollama/llama.cpp are fine in 2026; random x86 Windows apps are not.
NVIDIA DGX Spark — $9,449 CAD, configured in Quebec
When Strix Halo is the right buy
- You need Windows, or existing x86 tools.
- You want 128 GB unified memory without NVIDIA’s stack.
- You accept that many CUDA jobs will be slower or unsupported.
We source a current 128 GB Max+ 395 system (Framework Desktop class or equivalent). We will name the chassis on the order. We will not take money for AMD’s US-retail Halo SKU.
AMD Strix Halo 128 GB — $7,349 CAD
When neither is right
If the model fits in 24–48 GB of GDDR and you care about tokens/s, a discrete GPU usually wins bandwidth. That is a custom GPU inference rig — we used to build GPU mining rigs; PSU, PCIe, and exhaust are the same problem. See Spark vs custom GPU workstation and used 3090 for LLMs.
Canadian notes
Both boxes run locally. That removes a US API hop; it does not automatically satisfy Quebec Law 25 or PIPEDA, and it does not defeat CLOUD Act if a US-controlled operator still holds the keys. Ownership, admins, logs, and exit still need a written boundary — on-premises deployment, CLOUD Act, Law 25.
We are not the cheapest Spark listing in Canada. ExtremePC had units around $7,795 CAD on this date. Our number includes setup. If that is the wrong trade, say so — we would rather lose the cart than lie about the bill of materials.
Frequently asked questions
Is Strix Halo the same as Ryzen AI Halo?
Strix Halo is the APU. “Ryzen AI Halo” is AMD’s branded mini-PC sold through Micro Center in the US. Same silicon family; different SKU. We sell a sourcable 128 GB Max+ 395 desktop, not the US exclusive.
Can I dual-Spark for 256 GB?
NVIDIA documents connecting two Sparks for a larger unified pool. That is a quoted dual-unit job, not a second product on this page. Power is still two desk circuits, not a rack.
Which local AI computer (Canada) · Spark $9,449 · Strix Halo $7,349