Ryzen AI Max+ PRO 395 for Local LLMs:Strix Halo Measured Against a DGX Spark
On 2026-09-27 we ran one benchmark suite on two very different boxes: a Ryzen AI Max+ PRO 395 laptop (the chip the industry calls Strix Halo) and a DGX Spark class (GB10) unit. Identical llama.cpp builds, model files verified by SHA-256, one day, one operator. The Spark numbers live on our DGX Spark deep dive; this page is the AMD side of the ledger, including the software findings (Vulkan vs ROCm, the Windows shared-memory trap) that only show up when you run hardware for a living.

Strix Halo vs DGX Spark: What We Measured on the Same Build
Both machines ran llama.cpp build b11213, the same two model files (verified byte for byte by SHA-256), and the same benchmark commands. The table shows the winner in every cell, measured, not estimated. "GB10" is the DGX Spark class of machine; our unit is a Spark, and the MSI EdgeXpert units we own behave the same on the models below.
| Model and workload (same build, same files) | Strix Halo (Ryzen AI Max+ PRO 395, best backend) | DGX Spark class (GB10, CUDA) | Winner |
|---|---|---|---|
| Qwen3.6-35B-A3B Q4_K_M, prompt read (pp512) at 0 tokens of context | 1,313 tok/s (Vulkan) | 2,323 tok/s | Spark, 1.8x |
| Qwen3.6-35B-A3B, prompt read at 32K context | 760 tok/s (ROCm) | 1,907 tok/s | Spark, 2.5x |
| Qwen3.6-35B-A3B, generation at 0 tokens of context | 59.4 tok/s (Vulkan) | 69.5 tok/s | Spark, 1.17x |
| Qwen3.6-35B-A3B, generation at 32K context | 49.4 tok/s (Vulkan) | 58.5 tok/s | Spark, 1.18x |
| gpt-oss-120b MXFP4, prompt read at 0 tokens of context | 685 tok/s (ROCm) | 1,868 tok/s | Spark, 2.7x |
| gpt-oss-120b, generation at 0 tokens of context | 44.7 tok/s (Vulkan) | 59.5 tok/s | Spark, 1.33x |
| gpt-oss-120b, prompt read at 32K context | 359 tok/s (ROCm) | 1,381 tok/s | Spark, 3.85x |
| Dense Ollama models (qwen3.6:27b, gemma4:31b, qwen2.5:72b), generation | 8.8 to 11.4 tok/s | 10.5 to 12.6 tok/s | Spark, 11 to 19% |
| MoE Ollama models (qwen3.6:35b, gemma4:26b), generation | 51.8 to 54.4 tok/s | 66.1 to 74.5 tok/s | Spark, 22 to 28% |
| Measured memory bandwidth (torch triad probe) | 231 GB/s | 235 GB/s | Statistically tied |
Every number above is read from our stored result files from the 2026-09-27 run. The memory bandwidth result matters: the two machines have nearly identical bandwidth, so the decode gap on dense models (11 to 19 percent) is close to the compute and software gap, while the prompt-processing gap comes mostly from the Spark's CUDA path.
What Strix Halo Is: 128 GB of Unified Memory on an x86 Laptop Chip
Strix Halo is the industry nickname for AMD's Ryzen AI Max family, and the top SKU is the Ryzen AI Max+ PRO 395: 16 Zen 5 CPU cores, a Radeon 8060S GPU with 40 RDNA 3.5 compute units, and a single pool of up to 128 GB of LPDDR5x memory shared by the CPU and GPU. AMD rates the memory subsystem at 256 GB/s on an eight-channel LPDDR5x-8000 interface (AMD product page). Our torch bandwidth probe measured 231 to 235 GB/s in practice, which is 91 to 92 percent of the rated figure and nearly identical to the 235 GB/s we measured on the DGX Spark (95 percent of its rated 273 GB/s).
That last point deserves emphasis because it shapes every speed number on this page. Decode speed for a local LLM is mostly a memory-bandwidth problem: every generated token requires reading the model's active weights once. Two machines with the same measured bandwidth will land within a few percent of each other on decode, no matter what the marketing says. That is exactly what we measured: 59.4 versus 69.5 tokens per second on Qwen3.6-35B-A3B, with the gap explained by the Spark's more mature inference software as much as by its silicon.
Where Strix Halo genuinely differs is form factor and memory ceiling per dollar. The machine we tested is a laptop (an HP ZBook Ultra G1a) that also functions as a daily Windows 11 workstation. The DGX Spark class of machine is a small desktop appliance that does one thing. If your budget reaches exactly one box and you need it to also compile code, run your browser, join meetings, and run a 63 GB model, the Strix Halo machine does all of it and the Spark does not.
The Windows memory split, and why it matters
On Windows, the Ryzen AI Max+ PRO 395 does not hand all 128 GB to the GPU by default. Our test laptop exposed 95.8 GB of memory to Windows with 32 GB configured as dedicated graphics memory; the rest is shared, and the shared path is much slower once a model plus its context stops fitting in the dedicated slice. This setting lives in the BIOS and can be raised (the platform allows up to 96 GB dedicated), but raising it takes memory away from Windows. We measured with the 32 GB dedicated setting because that is how the machine arrived, and we flag the one workload where that setting visibly hurt (below). Linux distributions handle the split differently, and the newer 192 GB sibling chip (the Ryzen AI Max+ PRO 495, which we cover in our Gorgon Halo deep dive) exposes a larger GPU allocation.
Vulkan vs ROCm on Strix Halo: Both Are Usable, Neither Matches CUDA
We ran the same llama.cpp build twice on the Strix Halo laptop, once on its Vulkan backend (AMD's proprietary driver with cooperative-matrix support) and once on its ROCm/HIP backend (AMD's ROCm 7.13 wheels for the gfx1151 target). The results were close, with instructive differences:
| Workload | Vulkan | ROCm | Which to use |
|---|---|---|---|
| Qwen3.6-35B, generation at 0 / 8K / 32K | 59.4 / 56.4 / 49.4 tok/s | 54.3 / 50.0 / 44.7 tok/s | Vulkan, 8 to 11% faster at decode |
| Qwen3.6-35B, prompt read at 0 / 8K / 32K | 1,313 / 1,028 / 638 tok/s | 1,200 / 1,049 / 760 tok/s | Vulkan at short context, ROCm at 32K (+19%) |
| gpt-oss-120b, generation at 0 / 8K / 32K | 44.7 / 40.7 / 4.7 tok/s | 40.4 / 37.5 / 32.3 tok/s | ROCm: Vulkan collapses at 32K |
| gpt-oss-120b, prompt read at 0 / 8K / 32K | 628 / 450 / 92 tok/s | 685 / 587 / 359 tok/s | ROCm, decisively at long context |
Three findings from this table. First, generation speed favors Vulkan by 8 to 11 percent on the MoE model, so for chat-style workloads at short context, the Vulkan build of llama.cpp is the faster default. Second, prompt processing at long context favors ROCm, most dramatically on the largest model: at 32K tokens of context, ROCm read prompts at 359 tokens per second where Vulkan collapsed to 92 (and generated at 4.7 tokens per second, which is not usable). Third, the Vulkan collapse is reproducible and we believe we understand it: most of the 63 GB model sits in Windows shared memory with the 32 GB dedicated setting, and the Vulkan path could not keep the model plus 32K of KV cache resident. We did not test a larger dedicated setting because changing it requires a BIOS change, which was out of scope for that run.
The practical guidance: on Windows with the default memory split, use ROCm for any workload that involves long prompts on large models, and Vulkan for shorter-context chat and coding on mid-size models. On Linux, or after raising the dedicated allocation, re-test; the balance can shift. This backend sensitivity is a real cost of the AMD software ecosystem compared with CUDA, where one backend is simply the backend. It is the main reason our DGX Spark numbers have fewer footnotes than this page.
Ollama on Strix Halo: the friendly option, measured
Ollama 0.34.0 with its ROCm backend ran the same five-model set we ran on the Spark, and the picture for a home or small-office user is friendlier than the raw llama.cpp gap suggests. On the mixture-of-experts models people actually serve (qwen3.6:35b, gemma4:26b), Strix Halo delivered 51.8 to 54.4 tokens per second where the Spark delivered 66.1 to 74.5: a 22 to 28 percent gap that is perfectly livable for one user. On dense models (qwen3.6:27b, gemma4:31b, qwen2.5:72b at 4-bit), the gap narrowed to 11 to 19 percent, close to the measured bandwidth difference, because dense decode is nearly pure memory traffic.
Prompt reading is where the Spark pulls away in Ollama too: on the same 4,300-token prompts, the Spark prefilled qwen3.6:35b at 2,333 tokens per second against 980 on Strix Halo, and qwen2.5:72b at 379 against 98. If your workload is "paste a long document and ask questions," the Spark finishes reading two to four times sooner. If it is "chat with a model that fits," both machines feel similar, with the Spark modestly ahead.
One caution we hit during the comparison: published benchmarks of these platforms age badly in both directions. Our own April numbers on the same laptop class were 29 tokens per second on qwen3.6:35b; by September, after software updates, the same model ran 54.4. The Spark's equivalent row moved from 56 to 74.5. Any Strix Halo or Spark number older than a few months understates both platforms.
What Fits in 128 GB on Strix Halo (and What Does Not)
With 128 GB of unified memory and a BIOS adjustable graphics allocation, the practical model sizes are:
- Comfortable, fast: every MoE model up to about 40B parameters at 4-bit (Qwen3.6-35B-A3B at 22 GB, Gemma 4 26B-A4B, Nemotron 3.5 Lightning class), dense models to 32B at 4-bit, and gpt-oss-120b in MXFP4 at 63 GB. All of these ran at full speed from our tables above, with room for long context beside them.
- Fits with care: a 70B dense model at 4-bit (about 40 GB) leaves ample room; qwen2.5:72b ran at 4.0 tokens per second in Ollama, which is dense-model arithmetic on 231 GB/s, not a Strix Halo defect. The same model ran 4.6 on the Spark.
- Fit but fragile on Windows with the 32 GB dedicated split: gpt-oss-120b at 63 GB plus 32K of context tipped out of the fast path (the Vulkan collapse above). The same model was stable on ROCm and stable on both backends at 8K. If 120B-class models with long context are your target, either use ROCm, raise the dedicated allocation in BIOS, or move to Linux.
- Does not fit: full GLM 5.3 Flash, DeepSeek V4.1 Flash, or any model whose 4-bit file exceeds roughly 100 GB with context room left over. Those are the reason the 192 GB Ryzen AI Max+ PRO 495 exists (see our Gorgon Halo page) and the reason clusters of GB10 units exist (see our DGX Spark page).
Context capacity follows the same arithmetic. A 22 GB MoE model leaves more than 100 GB for KV cache, which is enough for multiple concurrent long-context sessions on Linux. The constraint on Windows is the split, not the chip: the same hardware with 96 GB dedicated holds a 63 GB model plus six-figure context comfortably.
When to Choose Strix Halo Over a DGX Spark (and When Not)
We publish the verdicts with the reasoning, including the cases where the answer is no.
| If your priority is... | Choose | Why, from our measurements |
|---|---|---|
| One box that is also your daily computer | Strix Halo | It is a complete x86 Windows or Linux laptop or mini PC. The Spark is an appliance. Our test machine ran the full benchmark suite and a normal Windows workload on the same day. |
| The lowest cost way to 128 GB of model memory | Strix Halo | A Strix Halo laptop or mini PC costs meaningfully less than a Spark class appliance (we make no price claims here; pricing is a sales conversation). The memory capacity is the same 128 GB. |
| Fastest decode for one user | Either, slight edge to the Spark | 17 to 33 percent on MoE models, 11 to 19 percent on dense models, same build. Both are in the "feels instant for one reader" range above 40 tok/s. |
| Reading long prompts fast (agents, document QA, RAG) | DGX Spark | Prompt processing was 1.8 to 3.85 times faster on the Spark in the same build. This is the gap we would pay for: an agent that re-reads a large context every turn spends most of its wall time in prefill. |
| Many users or agent fleets | DGX Spark, or a cluster | The CUDA software stack (vLLM, SGLang) carries concurrent load far better than anything on the AMD side today, and GB10 units cluster over 200 Gb links. Our GB10 vs H200 page shows where the ladder goes. |
| CUDA or NVFP4 software | DGX Spark | Strix Halo has no CUDA. ROCm and Vulkan cover llama.cpp and PyTorch well, but the newest inference engines land on CUDA first. |
| Silence and desk footprint | Strix Halo (laptop or fanless-adjacent mini PC) | A laptop on its own charger measured no wall power in our PDU because it barely draws any compared with the 130 to 158 W the Spark drew while generating. We did not meter the laptop, so we state the comparison qualitatively. |
| Models above ~100 GB at 4-bit | Neither: 192 GB Halo or a cluster | 128 GB is the ceiling for both. The Ryzen AI Max+ PRO 495 takes it to 192 GB; GB10 clusters take it further with better prefill. |
The one-sentence version: Strix Halo wins on cost per gigabyte of model memory, generality, and quiet; the Spark wins on prompt speed, software maturity, and concurrency. Both are excellent one-person machines, and neither is the right answer for a many-user deployment, which is a data-center tier question we answer on our LLM benchmark hub.
Strix Halo for Local LLMs: Frequently Asked Questions
Is the Ryzen AI Max+ PRO 395 good for running local LLMs?
Yes, for one to a few users. With 128 GB of unified memory it holds every model up to roughly 100 GB at 4-bit weights, and our measured decode speeds (51 to 59 tokens per second on 35B MoE models, 8 to 11 on dense 27B to 72B models) are usable for chat and coding. Its weakness is prompt reading speed, which measured 1.8 to 3.85 times slower than a DGX Spark in our same-build comparison.
Strix Halo or DGX Spark for a local LLM?
The Spark generated 17 to 33 percent faster on MoE models and read prompts much faster; Strix Halo costs less for the same 128 GB, is a full general-purpose computer, and is quieter. If your work is long-prompt agents, choose the Spark. If it is one-person chat and coding on models that fit, Strix Halo is the better value.
Should I use Vulkan or ROCm on Strix Halo?
Both are usable in llama.cpp. Vulkan was 8 to 11 percent faster at generation on our MoE tests; ROCm held up better on long prompts with the largest model we tested, where the Vulkan path collapsed under the Windows 32 GB dedicated-memory split. Start with Vulkan for chat, switch to ROCm for long-context work on big models.
Can Strix Halo run a 120B model?
Yes: gpt-oss-120b in MXFP4 (63 GB) ran at 40 to 45 tokens per second for generation on our laptop. With the Windows default 32 GB dedicated graphics allocation it slowed badly at 32K of context on Vulkan; ROCm stayed stable at 32.3 tokens per second. Raise the dedicated allocation in BIOS or use Linux for long-context 120B work.
How much memory bandwidth does Strix Halo have?
AMD rates the eight-channel LPDDR5x-8000 subsystem at 256 GB/s. Our torch triad probe measured 231 GB/s in practice, within a few percent of the 235 GB/s we measured on a DGX Spark. Bandwidth, not compute, sets decode speed, which is why the two machines are close on generation.
Can Strix Halo run Ollama?
Yes, and it is the friendliest path: Ollama 0.34.0 with the ROCm backend ran our five-model set at 4.0 to 54.4 tokens per second depending on the model. Mixture-of-experts models in the 26B to 35B range run 51 to 54 tokens per second, within about a quarter of the Spark.
What models do NOT fit in 128 GB?
Full-size GLM 5.3 Flash and DeepSeek V4.1 Flash, and any model whose 4-bit weights plus context exceed roughly 110 GB. For those you need the 192 GB Ryzen AI Max+ PRO 495 (Gorgon Halo) or a GB10 cluster.
Is Strix Halo better than a Mac Studio?
They trade places. Mac Studio M5 Max machines have more bandwidth (460 to 614 GB/s) and decode faster per box; Strix Halo reaches 128 GB of memory in a cheaper, general-purpose x86 machine and runs Ollama and llama.cpp on Linux natively. Our Mac Studio deep dive has the same-model numbers.
Can a Strix Halo machine do CMMC or compliance work?
The hardware can hold the models, and a single-user compliance assistant is exactly the workload unified memory suits. The compliance obligations live in your enclave design, not the chip; our CMMC compliance services team designs those. For retrieval-heavy compliance question answering, model choice matters more than the box, and our CMMC RAG scores are on the best local LLM page.
Does Strix Halo support CUDA or NVFP4?
No. CUDA is NVIDIA-only. Strix Halo runs ROCm/HIP and Vulkan; recent llama.cpp builds run well on both, and PyTorch has a ROCm backend. NVFP4 checkpoints are a NVIDIA format. If your stack depends on CUDA-only libraries, the Spark or an RTX card is the safer choice.
More Hardware Deep Dives From Our Benchmark Lab
Every page below uses the same prompts, the same quality gate and the same methodology, so the numbers line up across machines.
Sources and References
- AMD Ryzen AI Max+ 395 product page (16 Zen 5 cores, Radeon 8060S, LPDDR5x-8000, eight-channel 256 GB/s memory subsystem).
- NVIDIA DGX Spark page (GB10 Grace Blackwell Superchip, 128 GB unified memory).
- Meet the next AMD Ryzen AI Halo, AMD (the 192 GB sibling platform and its 160 GB GPU allocation).
- llama.cpp, the inference engine used for the same-build comparison (build b11213 on both machines).
- Ollama, version 0.34.0 on both machines for the five-model comparison.
- Qwen3.6-35B-A3B GGUF and the gpt-oss-120b GGUF, the model files verified by SHA-256 on both machines, Hugging Face.
- LLM benchmark methodology and the benchmark hub, Petronella Technology Group, Inc.
- DGX Spark vs Strix Halo vs Gorgon Halo: Local LLM benchmarks, our earlier-generation comparison.
- AMD Strix Halo AI workstations, our platform overview page for the product line.
Run the Right Model on the Right Unified-Memory Box
We benchmark the models you plan to run on the hardware you are considering, at your context lengths, before you buy. Our team builds private AI systems that keep sensitive data on equipment you control, from single-user Strix Halo desks to clustered GB10 deployments.
Prefer to read first? Explore our AI services or our on-premises AI practice.
Measured on the Petronella Technology Group, Inc. fleet, September to November 2026. Every number on this page traces to a stored result file with the engine build, model file hashes and backend recorded.