DGX Spark vs AMD Strix Halo: The $2,700 Question
A hands-on benchmark reveals NVIDIA's DGX Spark is 6x faster on prompt processing, but generation speed is a tie with a $2,000 AMD box. Here's what that means for your buy.
Audio Version
Everyone says the NVIDIA DGX Spark is "2-5x faster" than the $2,000 AMD Strix Halo box. A benchmark by KyzoroX on X [Link] tested both machines with the same 30B parameter model and found something the marketing materials don't tell you: it's 6x faster on prompt processing, and near-identical on generation.
Which means most people are about to overpay by $2,700.
Key Benchmark Results — 30B Model, 128GB Unified Memory
Read that twice. On the number you actually watch happen — text appearing on screen — a $2,000 box ties a $4,700 one. The Spark's entire premium lives in one place: prefill.
The Two Phases of LLM Inference
To understand why these numbers matter, you need to understand the two distinct phases of running an LLM locally.
Prefill: Reading Your Context
When you send a prompt to an LLM, the model first processes every token in your input. This is the prefill phase — the model reads your entire context window, builds the key-value cache, and prepares to generate. This phase is compute-bound: it's limited by how many floating-point operations your GPU can perform per second.
NVIDIA's Blackwell Tensor Cores absolutely demolish AMD's RDNA 3.5 iGPU here. The DGX Spark's 5th-gen Tensor Cores with FP4 support deliver 1 PetaFLOP of AI compute. That's why it processes prompts at 2,107 tokens per second versus AMD's 342 — a 6x advantage that's pure silicon architecture.
Decode: Writing the Answer
Once prefill is done, the model generates tokens one at a time. This is the decode phase — the speed you actually feel as a user. Each token generation requires loading the model weights from memory, computing attention, and producing one output token. This phase is bandwidth-bound: it's limited by how fast you can move data from memory to the compute units.
Here's where the story changes. Both systems have roughly similar memory bandwidth — the DGX Spark's 128GB LPDDR5x delivers ~273 GB/s, while the Strix Halo's LPDDR5X-8000 hits ~256 GB/s. They're close enough that the decode speeds are nearly identical: 84 t/s versus 73 t/s. That's a 15% difference — statistically meaningful, but not something you'd notice watching text stream on your screen.
Under the Hood: Architecture Comparison
NVIDIA DGX Spark — The Blackwell Desktop
| Spec | Details |
|---|---|
| Chip | NVIDIA GB10 Grace Blackwell Superchip (dual-die, NVLink-C2C) |
| GPU | Blackwell, 48 SMs, up to 2.55 GHz |
| CPU | 20 ARM Cortex cores (10× X925 + 10× A725) |
| Tensor Cores | 5th-gen with FP4 support |
| AI Performance | 1 PFLOP (FP4 sparse) / 31 TFLOPS FP32 / 1,000 TOPS |
| Memory | 128GB LPDDR5x unified, 273-301 GB/s |
| Storage | 4TB SSD (base) |
| TDP | ~300W peak |
| Price | $4,699 (raised from $3,999 in Feb 2026) |
| Software | CUDA, TensorRT-LLM, vLLM, llama.cpp CUDA backend |
AMD Strix Halo (Ryzen AI Max+ 395) — The Unified APU
| Spec | Details |
|---|---|
| Chip | AMD Strix Halo APU (single monolithic, 2 CCD chiplets) |
| CPU | 16 Zen 5 cores / 32 threads, 3.0 GHz base / 5.1 GHz boost |
| GPU | Radeon 8060S iGPU — 40 RDNA 3.5 CUs (gfx1151) |
| NPU | XDNA 2, 50 TOPS AI acceleration |
| Memory | Up to 128GB LPDDR5X-8000 unified, 256 GB/s |
| TDP | 45W–120W configurable |
| Price | $1,499–$2,299 (third-party mini PCs with 128GB) |
| Software | ROCm 7.0 (Windows), Vulkan backend in llama.cpp |
The Software Reality Check
The benchmarks tell only half the story. The software ecosystem is where the real-world experience diverges sharply.
CUDA: The 18-Year Head Start
NVIDIA's CUDA ecosystem is the default for AI development. Every major framework — PyTorch, TensorFlow, JAX, vLLM, TensorRT-LLM — supports CUDA with day-one reliability. When a new model ships, CUDA is the first backend tested. When something breaks, there are millions of Stack Overflow answers, GitHub issues, and community tutorials to help you fix it.
On DGX Spark, you get the full NVIDIA stack: CUDA, TensorRT-LLM for optimized inference, vLLM for high-throughput serving, and llama.cpp with CUDA backend. New models and quantization formats (FP4, NVFP4, MXFP4) land on CUDA first. The experience is production-ready out of the box.
ROCm: Improving Fast, Still Catching Up
AMD's ROCm 7.0 landed in September 2025 with native Windows support and day-zero PyTorch integration — a significant milestone. But here's the catch: ROCm doesn't officially support Strix Halo's gfx1151 GPU yet. iGPU and APU users primarily rely on the Vulkan backend in llama.cpp.
Community reports suggest the Vulkan backend actually outperforms ROCm on Strix Halo for token generation, while ROCm is faster for prompt processing when it works. ROCm 7.2.1 improved AMD support, but the Vulkan path remains stuck on older llama.cpp code from late 2025.
The practical implication: when a new AI tool ships CUDA-only, you're debugging ROCm compatibility at midnight. The memory on Strix Halo is real — 128GB unified is glorious — but the software lottery is too.
Who Should Buy What
So the buy is simple once you know your bottleneck:
AMD Strix Halo (~$2,000)
- You chat, draft, and code with short prompts. Same generation feel as DGX Spark, pocket $2,700.
- You want Windows compatibility. Native Windows 11 support, plus gaming on the side.
- You value cost-per-token. 2.3x better value for everyday inference workloads.
- You're comfortable with Vulkan/ROCm. Willing to troubleshoot when new tools ship CUDA-only.
NVIDIA DGX Spark ($4,699)
- Your work is huge context. Long agent runs, RAG over hundreds of docs, giant files — that 6x prefill is the Spark earning its price.
- You need production reliability. CUDA just works. Every model, every framework, day one.
- You do fine-tuning or training. TensorRT-LLM, full CUDA stack, FP4 quantization support.
- You can't afford software friction. Your time is worth more than the $2,700 premium.
The Bigger Picture
This benchmark reveals a fundamental truth about local AI hardware: buy for your bottleneck, not the biggest number on the box.
NVIDIA's marketing emphasizes the 6x prefill advantage — and it's real. But for most users, prefill happens once per conversation. Decode happens for every single token the model generates. If you're chatting with an LLM for 30 minutes, you'll spend maybe 2 seconds in prefill and 1,800 seconds in decode. The DGX Spark's advantage is concentrated in the phase you barely notice.
The Strix Halo's value proposition is compelling for a specific class of users: developers who primarily run conversational AI workloads, want native Windows support, and don't mind occasional software friction. At $2,000 versus $4,699, it delivers nearly identical generation speed for less than half the price.
But if your workflow involves massive context windows — running autonomous agents, processing hundreds of documents through RAG, or working with giant codebases — that 6x prefill advantage compounds quickly. What takes 30 seconds on Strix Halo takes 5 seconds on DGX Spark. Over a day of heavy use, those seconds add up to hours.
Nothing else in the compact local AI space touches the DGX Spark on prefill. But nothing else comes close to the Strix Halo on price-to-performance for generation-bound workloads. The right choice depends entirely on what you're actually doing with the machine.