Strix Halo vs DGX Spark: the 128-machine showdown GB

#Two 128 GB systems, two philosophies

The Ryzen AI Max+ 395 is an x86 APU: a Zen 5 CPU (16 cores), a Radeon 8060S iGPU (RDNA 3.5, 40 CUs), and an XDNA 2 NPU, all sharing up to 128 GB of LPDDR5X over a 256-bit bus. It's available in mini PCs (Framework Desktop, GMKtec EVO-X2, Beelink GTR9) and a few laptops. It's a mainstream platform, running Windows or Linux, that runs Ollama and LM Studio like any PC.

The DGX Spark is a different beast: a GB10 chip (Grace Blackwell architecture) with a 20-core ARM CPU and a Blackwell GPU with Tensor Cores, also paired with 128 GB of unified LPDDR5X. It's a NVIDIA-first device, shipped with DGX OS (a customized Ubuntu) and the full CUDA stack. Announced official price: $4,699.

#1. Specs side by side

The most surprising point at first glance: memory bandwidth is nearly identical. Yet it is what governs token generation speed. This explains why the two machines perform similarly in chat.

Memory

128 GB of unified LPDDR5X shared by both. Enough for a 70B Q4 (~40 GB) plus context, or an MoE such as Qwen3-235B with aggressive quantization.

Bandwidth

Strix Halo: ~256 GB/s theoretical (256-bit, LPDDR5X-8000). DGX Spark: ~273 GB/s. The real-world difference is negligible for generation.

GPU compute

Radeon 8060S RDNA 3.5, 40 CU, with no dedicated Tensor Cores. On the Spark side: a Blackwell GPU with Tensor Cores and native FP4 support — an order of magnitude above it in raw compute.

CPU

AMD: Zen 5, 16 x86 cores. NVIDIA: ARM Grace, 20 cores. x86 remains simpler for mainstream software.

OS

Strix Halo: Windows 11 or Linux, your choice. DGX Spark: DGX OS (Ubuntu) with drivers and CUDA preinstalled.

Power consumption

Strix Halo: 55–120 W power envelope depending on the chassis. DGX Spark: ~240 W at the wall under load, with a dedicated power supply.

#2. Tokens/sec: the close match

For pure generation (decode), both machines are limited by memory bandwidth, not compute. The result: similar figures, with Spark holding a slight advantage thanks to its marginally higher bandwidth and highly optimized CUDA kernels. Here are representative measurements published by the community in 2026, using equivalent Q4 quantization and a short prompt.

Qwen3-30B-A3B (MoE) — Strix Halo

~100 tok/s during generation. The MoE activates only 3B parameters, which explains the high speed despite its size.

Qwen3-30B-A3B (MoE) — DGX Spark

~100-115 tok/s. Small difference: both are memory-bound on this lightweight MoE.

Dense 70B Q4 model — Strix Halo

~4–5 tok/s. A dense 70B saturates the bandwidth; the experience remains usable but slow.

Dense 70B Q4 model — DGX Spark

~5-6 tok/s. Same here, memory-bound. The difference is measured in fractions of a token/s.

8B Q4 model—the two

50–70 tok/s, more than comfortable for interactive chat.

The conclusion of this section is counterintuitive: if you look only at tokens per second in chat, the two machines are nearly equivalent, and the much cheaper Strix Halo seems to win. But that number tells only half the story.

#3. Prompt processing: the real difference

Prompt processing (or prefill) is the phase in which the model reads and encodes your input prompt before generating the first token. Unlike generation, this phase is compute-bound, not memory-bound. This is where the DGX Spark's Tensor Cores make all the difference.

Strix Halo — prefill

~340 tok/s of prompt processing on a typical 30B model. The RDNA 3.5 iGPU has no dedicated matrix units, so it hits its ceiling quickly.

DGX Spark — prefill

about 5x faster on the same models, thanks to Blackwell Tensor Cores and FP4 support. With large contexts, the gap can widen further.

Why does it matter? Because as soon as you move beyond short chats, prefill dominates perceived response time. A RAG system injecting 8,000 tokens of context, an agent rereading its entire history on every turn, long-document analysis, batch processing: in all these cases, you wait for the prefill before seeing anything.

Concrete example: a 4,000-token prompt. At 340 tok/s, the Strix Halo takes ~12 seconds just for prefill before it starts responding. The DGX Spark, ~5x faster, completes the same phase in 2–3 seconds. In an intensive RAG session with dozens of queries, this difference becomes the dominant part of the experience.

#4. Price and availability

This is where the Strix Halo pulls ahead—and by a wide margin. Its price-to-memory ratio is its knockout argument.

DGX Spark

$4,699, official NVIDIA price. Available through NVIDIA and partners (Dell, Asus, HP, Lenovo). Positioned for professionals and developers.

Strix Halo — mini PC

Framework Desktop, GMKtec EVO-X2, Beelink GTR9: with a 128 GB configuration, you generally land between ~$1,700 and ~$2,500. About half the price of the Spark for the same memory.

Strix Halo — laptop

A few laptops (e.g., HP ZBook Ultra and Asus ROG Flow Z13) include the APU, a rare advantage for running a 128 GB LLM on the go.

Availability

Strix Halo has been openly available from several system builders since mid-2026. The DGX Spark follows a more constrained NVIDIA schedule depending on the region.

#5. Software ecosystem

Beyond the hardware, the software stack has a major impact on day-to-day usability—and that's where NVIDIA benefits from years of CUDA head start.

DGX Spark — CUDA

The entire NVIDIA ecosystem works natively: vLLM, TensorRT-LLM, PyTorch, and NGC containers. For fine-tuning, batch serving, or research, this is the best-documented platform.

Strix Halo — ROCm / Vulkan

Inference works well through Ollama and llama.cpp (Vulkan or ROCm backend). Fine-tuning and advanced frameworks remain more hands-on on RDNA 3.5 than on CUDA.

Consumer-friendly simplicity

Strix Halo's advantage for the average user: Windows, Ollama with one command, LM Studio with one click. Spark assumes you are comfortable with Linux/Ubuntu.

Multi-client serving

DGX Spark advantage: vLLM and TensorRT-LLM provide much higher batching and aggregate throughput for serving a team or an app.

#6. Which profile for which machine

1. 01
   You mainly do local chat and MoE

   Strix Halo. Generation tokens/sec are on par with Spark, the price is half as much, and Windows + Ollama makes it trivial to use. A Qwen3-30B-A3B at ~100 t/s is an excellent daily driver.
2. 02
   You're building a RAG system or long-context agents

   DGX Spark. 5x faster prompt processing transforms the experience as soon as you repeatedly inject large contexts. This is the scenario where the additional price is justified.
3. 03
   You want to fine-tune or do research

   DGX Spark. The CUDA stack (PyTorch, TensorRT-LLM, NGC containers) is incomparably better equipped than ROCm on an APU for training.
4. 04
   You're looking for the best memory-to-euro ratio

   Strix Halo, without hesitation. 128 GB unified memory (prices vary widely, so verify), generally cheaper than the Spark, with the x86 versatility of a true desktop PC.
5. 05
   You want 128 GB on the go

   Strix Halo, via a laptop equipped with the APU — a category the DGX Spark, a desktop device, does not cover.

#Verdict

Deliberate verdict: a tie. This duel has no single winner, and that's not a cop-out — it's the result of the measurements. Both chips offer 128 GB of unified memory and comparable generation speed; each wins half the match. Strix Halo wins on price-to-memory ratio, versatility (it's also an excellent PC), and interactive chat; GB10 wins on prompt processing, fine-tuning, clustering, and the CUDA ecosystem. Two philosophies, two winners — the only loser is the conventional GPU with limited VRAM.

The misleading number

~100 tok/s generation on both sides with Qwen3-30B → seems like a tie, with Strix Halo having the price advantage.

The deciding number

Prompt processing at ~340 tok/s (Strix Halo) versus ~5× more (DGX Spark). It’s what makes the difference as soon as you leave short chats.

The simple rule

Short context + tight budget → Strix Halo. Long context, agents, RAG, fine-tuning → DGX Spark.

#Frequently asked questions

#Go further

To dig deeper into each machine and the hardware context around this head-to-head:

Ryzen AI Max+ 395 (Strix Halo)

The dedicated guide to AMD APUs: available mini PCs and laptops, accessible 70B+ models, and real-world performance compared with dedicated GPUs and Macs.

NVIDIA DGX Spark

The 128 GB AI mini PC examined alongside conventional GPUs: performance benchmarks, use cases, and limitations.

Choose your quantization (Q4, Q5, Q8, FP16)

To understand why a 70B fits in 40 GB in Q4 and what each step down costs in quality on these unified-memory machines.