AMD Ryzen AI Halo vs NVIDIA DGX Spark 2026: Which 128GB AI Dev Kit Actually Pays Off

This article was originally published on runaihome.com

TL;DR: Both boxes carry 128GB of unified memory and both are bandwidth-bound, so token generation is nearly tied — ~34 tok/s (Halo) vs ~39 tok/s (Spark) on gpt-oss 120B. DGX Spark wins prompt processing by roughly 5× and brings CUDA; AMD's Halo gives 2TB of storage and Linux-first ROCm for the same $3,999. For pure inference, a cheaper Strix Halo box wins.

*NVIDIA raised the Founders Edition from $3,999 to $4,699 in 2026 as memory prices spiked.

Honest take: If you only run inference on MoE models, both are bandwidth-bound and roughly tied on the tokens you actually feel — so the cheaper Strix Halo box (like the GMKtec EVO-X2 at ~$1,999) is the smarter buy. Pay the DGX Spark premium only if you fine-tune, write CUDA, or run prefill-heavy agentic workloads.

The matchup isn't $3,999 vs $2,999

The headline that's been circulating — AMD's $3,999 dev kit "tackling" a $2,999 DGX Spark — quietly compares two different things. NVIDIA's own DGX Spark Founders Edition launched at $3,999 with a 4TB SSD, and as of 2026 NVIDIA raised it to $4,699 because of the same memory-price spike that's hitting DDR5 and SSD buyers. The $2,999 figure belongs to the ASUS Ascent GX10, an OEM variant of the DGX Spark with the same GB10 Grace Blackwell Superchip and 128GB of memory — but only 1TB of storage instead of 4TB.

So the real picture, dollar for dollar, is:

• AMD Ryzen AI Halo — $3,999, Ryzen AI Max+ 395, 128GB, 2TB SSD
• DGX Spark (ASUS GX10) — $2,999, GB10, 128GB, 1TB SSD
• DGX Spark (Founders) — $3,999 (now $4,699), GB10, 128GB, 4TB SSD

AMD slots its dev kit between the two NVIDIA configs on price, gives you double the storage of the cheap Spark, and matches the Founders Edition exactly at $3,999 before NVIDIA's hike. That's the framing AMD wants: same money, more SSD, and — they claim — leadership tokens-per-dollar.

What's actually in each box

The AMD Ryzen AI Halo Developer Platform is built on the Ryzen AI Max+ 395 — the same "Strix Halo" APU in the GMKtec EVO-X2 and the chip we tore down in our Strix Halo deep dive. It pairs 16 Zen 5 cores (3.0GHz base, 5.1GHz boost) with a Radeon 8060S iGPU (40 RDNA 3.5 compute units) and an XDNA 2 NPU rated at 50 TOPS, for a platform total AMD quotes at 126 TOPS. Memory is 128GB of LPDDR5X-8000 on a 256-bit bus — 256 GB/s — and the kit ships with a 2TB PCIe 4 SSD, 10GbE LAN, Wi-Fi 7, four USB-C ports, and an aluminum chassis the size of a paperback (149 × 149 × 43mm). Pre-orders run through Micro Center, with availability around July 10, 2026, and you pick Linux or Windows at no price difference — a tell that AMD is aiming this squarely at developers.

The NVIDIA DGX Spark is a different animal under the hood. Its GB10 Grace Blackwell Superchip glues a 20-core Arm CPU (10 Cortex-X925 + 10 Cortex-A725) to a Blackwell GPU that NVIDIA rates at up to 1 petaFLOP of sparse FP4 tensor performance. Memory is also 128GB of unified LPDDR5x, but at a slightly higher 273 GB/s. It runs DGX OS (Ubuntu-based) and, crucially, the full CUDA stack.

The spec sheets converge on the thing that matters most for local LLMs: 128GB of unified memory on both. That's enough to hold models no single 24GB consumer GPU can — gpt-oss 120B, Qwen3-235B at lower quant, dense 70B with room to spare. The question is how fast each one actually moves tokens through that memory.

Token generation: nearly a tie

Here's the result that surprises people. On token generation — the decode phase, where the model streams one token at a time and you watch words appear — the two boxes are within a few tokens per second of each other.

On gpt-oss 120B (the MoE model both vendors lean on for demos), independent testing puts the Ryzen AI Max+ 395 at 34.13 tok/s against the DGX Spark's 38.55 tok/s. That's a 13% NVIDIA lead, not a generational gap. The reason is simple and it's the same reason every box in this class behaves the way it does: decode is memory-bandwidth-bound, not compute-bound. The GB10's 273 GB/s and Strix Halo's 256 GB/s are within 7% of each other, so the tokens-per-second they can sustain on the same model are within 7% too. NVIDIA's enormous FP4 compute advantage simply doesn't get used during decode.

Per-model llama.cpp numbers fill in the rest of the picture on the DGX Spark side:

On the AMD side, gpt-oss 20B on the Ryzen AI Max+ 395 lands around 30–33 tok/s generation with roughly 400 tok/s prompt processing. Notice the pattern across both platforms: MoE models (gpt-oss, Qwen3-30B-A3B) fly because only a few billion parameters activate per token, while dense 70B craters to ~2.6 tok/s on the Spark — the same single-digit decode we measured in our GMKtec EVO-X2 review. Neither of these machines is a good dense-70B box. Both are MoE boxes that happen to have enough memory for big models.

For reference, human reading speed is about 7–10 tok/s, so anything in the 30–60 range feels comfortably interactive and the dense-70B ~2.6 tok/s is a "start it and walk away" experience on either platform.

Prompt processing: where NVIDIA earns its badge

The gap that actually separates these two boxes isn't decode — it's prefill, the prompt-processing phase that runs before the first token appears. On gpt-oss 120B, the DGX Spark processes the prompt at roughly 1,723 tok/s versus Strix Halo's 339.87 tok/s — about 5× faster. Prefill is compute-bound, and this is exactly where Blackwell's tensor cores and the mature CUDA kernels do their work while Strix Halo's RDNA 3.5 iGPU falls behind.

This matters more than the spec-sheet symmetry suggests, and it maps directly to your workload:

• Short prompts, chat, casual coding autocomplete → prefill is a rounding error. The two boxes feel identical.
• Long context: RAG over big documents, full-repo code analysis, 32K+ token agentic loops → prefill dominates time-to-first-token. A 5× prefill advantage is the difference between a 3-second wait and a 15-second one, every turn. Here the DGX Spark pulls clearly ahead.
• Fine-tuning / training → not a contest. CUDA, cuDNN, and the entire PyTorch training ecosystem run first-class on the Spark; a QLoRA pass on Llama 3.3 70B has been measured at over 5,000 tok/s throughput on the GB10. ROCm fine-tuning on Strix Halo works but is rougher and slower.

ROCm vs CUDA: the real tax

The benchmark you can't put in a table is software friction. NVIDIA is selling a decade of CUDA momentum — PyTorch wheels that just work, every inference runtime tested on it first, Docker images, NIM microservices, forum answers for every error. If your workflow touches custom CUDA kernels, vLLM tensor-parallel, NeMo,