DGX Spark vs Strix Halo vs Mac Studio for LLMs

Choose by model architecture, not by badge. If you run mixture-of-experts models, a GB10 box is the fastest of the three per dollar — published llama.cpp benchmarks put gpt-oss-120b at 60.6 tok/s on a DGX Spark, and the ASUS Ascent GX10 clone with the identical chip is $3,099.99. If you run dense 70B models, only the Mac is usable: Apple's 819 GB/s on the M3 Ultra is roughly 3x what either 128GB x86/Arm box offers, and LMSYS measured dense Llama 3.1 70B at 2.7 tok/s on the Spark. If you just want 128GB as cheaply as possible and can live with 256 GB/s, a Strix Halo mini-PC is the price floor at $1,999.99 list.

That is the whole decision, and it turns on one number per machine: 273 GB/s (NVIDIA GB10), 256 GB/s (AMD Ryzen AI Max+ 395), 819 GB/s (Apple M3 Ultra) — all three vendor-published. Token generation is memory-bandwidth-bound, so those figures set a ceiling that no driver update, quantisation trick or firmware revision moves.

We have not bought any of these three machines. Every performance figure below is a published third-party measurement and is attributed to whoever ran it, with the date and framework, because early numbers in this category scattered badly with software maturity. What we add is the arithmetic that connects the benchmarks to the bandwidth specs — arithmetic you can check — and a clear statement of what we could not verify. Our single-product deep dives on each machine are the DGX Spark review, the Strix Halo / AI Max+ 395 guide and the Apple M5 guide.

Which of the three is actually fast enough to use?

On paper the three are nearly interchangeable — 96-128GB of unified memory in a box you can carry. On bandwidth per gigabyte they are not even in the same category.

The "GB/s per GB" row is the one to sit with. The Mac moves each gigabyte of its memory pool roughly four times as fast as either 128GB box. That ratio is why the three machines feel completely different despite spec sheets that look similar, and it is why "which has more memory" is the wrong opening question.

Here is the arithmetic that turns bandwidth into a speed ceiling, because you should be able to redo it for any model you care about. Generation reads roughly the model's active weights once per token, so:

• A dense 70B at Q4 is about 40GB of weights. 273 / 40 = ~6.8 tok/s ceiling on GB10. 256 / 40 = ~6.4 tok/s on Strix Halo. 819 / 40 = ~20.5 tok/s on an M3 Ultra. Those are theoretical maxima; real numbers land below them.
• A dense 70B at FP8 is about 70GB. 273 / 70 = ~3.9 tok/s on GB10 — and LMSYS measured 2.7. The hardware is already near its ceiling. That is physics, not a driver bug.
• gpt-oss-120b is a mixture-of-experts model: 117B total parameters but only 5.1B active per token, in MXFP4, per OpenAI's model card. Only a few gigabytes move per token, which is how a 117B model outruns a 70B one by more than 20x on the same machine.

So the honest framing is not "fast box versus slow box." It is: all three are capacity machines with a bandwidth budget, and MoE models are the models built for that budget. The memory bandwidth explainer works through why the relationship holds so tightly.

How fast does each box run real models?

Every number below was published by a named third party. None of them is a controlled head-to-head — different models, runtimes, quantisations and dates — so read the pattern, not the deltas.

Three readings that matter more than any single cell:

1. The MoE column is where the Spark earns its price. 60.6 tok/s against ~45 is a real gap — about 35% — and it lines up with the 273-versus-256 GB/s spec difference plus a more mature CUDA runtime. Note the LM Studio row though: the same gpt-oss-120b ran at 31.41 tok/s on a Beelink GTR9 Pro versus ~45 tok/s in ServeTheHome's AMD dev-system testing. Same chip, same memory. Runtime choice moves these numbers more than brand choice does.

2. The dense 70B row is the whole argument for the Mac. 2.7 tok/s is slower than reading pace; LMSYS's own framing was that 70B and 120B-class dense models are best suited to prototyping rather than production on GB10. 15 tok/s is comfortable chat. And note the cross-check: the M5 Max runs at 614 GB/s, and 614 / 40 = 15.4 — the measured 15 tok/s sits almost exactly on the bandwidth ceiling, which is the strongest evidence available that this arithmetic describes reality.

3. Prefill is a genuine Spark advantage nobody talks about. Nearly 8,000 tok/s of prompt processing on an 8B model is Blackwell tensor compute doing work that bandwidth does not gate. If your day is long-context RAG ingestion, document summarisation or agent loops with fat prompts, that column matters more to you than decode does.

One honest calibration on all of the above: ServeTheHome's launch-day Spark review measured gpt-oss-120b at 14.5 tok/s, roughly a quarter of the llama.cpp maintainer's figure, and commenters flagged likely configuration problems. Early numbers in this category scattered enormously with software maturity. Whenever you see a tok/s figure for any of these three machines — including on this page — check the date, the framework and the quantisation before comparing it to anything.

How much memory can each one actually give a model?

Total memory and model-addressable memory are different numbers on two of the three machines, and the defaults will cost you a working setup if you do not change them.

The Strix Halo default is the single most common way people conclude the box is broken. Out of the crate it hands the iGPU something like 16-32GB and leaves the rest as system RAM, so a 70B model refuses to load on a 128GB machine. You set it in the BIOS: our Strix Halo guide recommends 96GB for LLM-only use (leaving 32GB for the system) and up to 110GB as the maximum practical allocation for a 70B at BF16 with long context. Do it before you conclude anything about performance.

On the Mac the cap is real but softer. macOS will not let Metal wire all of unified memory; the limit is exposed as recommendedMaxWorkingSetSize, and Apple's MLX documentation points at sudo sysctl iogpu.wired_limit_mb as the way to raise it — while warning that the wired limit "should remain strictly less than the total memory size." A 70-75% planning rule is the safe default, so budget about 67-72GB of the M3 Ultra's 96GB for weights plus KV cache. Raising the sysctl a few gigabytes is reasonable; raising it aggressively trades a clean OOM for system-wide memory pressure, and it does not survive a reboot.

The GB10 boxes genuinely have no split to manage, which is a small, real quality-of-life win. NVIDIA's spec sheet describes the 128GB LPDDR5X as coherent unified memory shared by the Grace CPU and the Blackwell GPU, and in practice gpt-oss-120b (about a 65GB download through Ollama) loads with substantial headroom.

Worth stating because it cuts against the whole category: capacity is not what separates these machines. All three hold 96-128GB. You are choosing on bandwidth, software stack and price — nothing else on the spec sheet is a differentiator.

What does each box refuse to run?

This is where the comparison stops being about numbers. Each of the three is a different software island, and the honest question is not "which is fastest" but "which one refuses to run the thing you depend on."

DGX Spark / GB10 clones — CUDA, but on Arm. The mainstream stack is solid: the standard Linux Ollama installer works on DGX OS (it is Ubuntu-based), llama.cpp builds with the ordinary -DGGML_CUDA=ON flags, and vLLM published an official DGX Spark deployment guide in June 2026. TensorRT and NIM containers are the reason to pick this box if your prototypes deploy to NVIDIA datacenter GPUs — no other machine here offers that continuity. What breaks: aarch64. A quantisation script, a prebuilt Python wheel or a Docker image that assumes x86 will still occasionally bite you, and the fix is compiling it yourself.

Strix Halo — ROCm on ordinary x86. ROCm 6.3 and newer support gfx1151; Ollama, llama.cpp and vLLM-ROCm all work, and Vulkan is a solid fallback that official Ollama now enables by default. Some workloads still want HSA_OVERRIDE_GFX_VERSION=11.5.1 set, because the iGPU reports as 11.5.1. What does not run: anything CUDA-only — TensorRT-LLM, ExLlamaV2, and the pile of fine-tuning tooling built on CUDA kernels. vLLM-ROCm functions but is throughput-bound by 256 GB/s rather than by the runtime. Our AMD ROCm setup guide covers the install path.

Mac Studio — Metal and MLX, and nothing from the NVIDIA world. llama.cpp's Metal backend and Ollama both work well, and MLX is genuinely good — the fastest published Apple numbers in the table above are MLX runs. What does not run: CUDA, at all. No TensorRT-LLM, no ExLlamaV2, no CUDA vLLM path, none of the bitsandbytes/xformers ecosystem. Fine-tuning means MLX-LoRA or a cloud GPU. If your workflow has a single CUDA-shaped dependency in it, the Mac is out regardless of how good its bandwidth is. Our Mac local AI setup guide covers what the Metal backend does and does not do well.

The compressed version: the Spark's unassailable niche is CUDA continuity, the Strix Halo's is being a normal x86 Linux machine that happens to have 128GB, and the Mac's is bandwidth you cannot buy anywhere else at this size. Nobody wins on software in the abstract; you win or lose on your own dependency list.

What do the three actually cost right now?

Prices below were checked in the first week of August 2026 at the named vendor stores, except the GB10 row, which comes from our mid-July 2026 market survey. This category moved violently through the memory shortage, so treat everything here as a snapshot and re-check before ordering.

Two things in that table are worth more than the rest of it.

First: buy the GB10 clone, not the badge. The ASUS Ascent GX10 carries the same GB10 Grace Blackwell chip and the same 128GB as the DGX Spark for $1,600 less than the Founders Edition, which NVIDIA raised from $3,999 to $4,699 in late February 2026 citing the DRAM/NAND shortage. Unless first-party support or the FE's 4TB drive specifically matters to you, the price difference is a badge tax.

Second: the cheap Strix Halo listings and the buyable Strix Halo boxes are different boxes. The EVO-X2's $1,999.99 is the best price per gigabyte anywhere in this comparison — and every configuration showed unavailable. The Beelink GTR9 Pro, which ServeTheHome's October 2025 review said often sold for "$1999 or less," listed at $4,349 ten months later with no hardware change. In this market, the cheapest acceptable box that is actually in stock wins, which in early August 2026 meant the $3,639 Minisforum MS-S1 Max. Our Strix Halo mini-PC roundup tracks the stock situation box by box, and the memory shortage explainer covers why the whole category inflated.

A note on the Mac column that surprises people: Apple's memory ceiling went down, not up. As of August 2026 the Mac Studio ships the M4 Max (36GB or 64GB) and the M3 Ultra, and the M3 Ultra is now fixed at 96GB after Apple removed the 512GB and then the 256GB tiers during the DRAM shortage, per MacRumors' reporting. There is no M5 Ultra. That is exactly why the 128GB MacBook Pro M5 Max ended up on this list at all — it is currently the only 128GB Mac Apple sells. If you want the full Apple picture, the Apple Silicon buying guide maps every configuration.

Which one should you buy for your model list?

Start from the models, not the hardware. The machine follows.

That last row is the one most people need and least want to hear. These are capacity machines. If your models fit on a 24GB card, a used RTX 3090 at 936 GB/s will beat all three of these boxes on speed for a fraction of the money — the bandwidth table above makes it obvious. The entire reason this category exists is models that a consumer GPU cannot hold at all.

If you are cross-shopping a Mac against a conventional GPU tower rather than against these appliances, that is a different comparison and we have it: Mac Studio vs a PC build.

What could we not verify?

Stating the gaps explicitly, because a comparison that hides them is worse than a shorter one.

• No controlled head-to-head exists. Every figure in the benchmark table comes from a different reviewer, on a different runtime, on a different date, on different models. We have not normalised them and neither has anyone else. Cross-machine deltas under about 20% should be treated as noise.
• We have no gpt-oss-120b number for the Mac Studio M3 Ultra. The bandwidth arithmetic suggests it should be strong, and we are not going to publish a tok/s figure we cannot attribute to someone who ran it. The Apple dense-70B numbers we do cite are from an M5 Max at 614 GB/s, not the M3 Ultra at 819 — treat 15 tok/s as a floor for the Mac Studio rather than a measurement of it.
• Nobody has published dual-Strix-Halo results we would stand behind. The question comes up constantly; the honest answer is that these boxes have no NVLink-class interconnect, the fastest links available on them are dual 10GbE or USB4, and we have not seen a credible two-node LLM benchmark. The Spark is the only machine here with a documented clustering story (200GbE RDMA, per ServeTheHome).
• The prices will be wrong before this page is old. One box in this comparison changed by more than 100% in ten months. The dated snapshot above is a starting point for your own checking, not a quote.
• Software is moving fast on two of the three. ROCm's gfx1151 support and NVIDIA's aarch64 stack are both improving on a monthly cadence, and NVIDIA has claimed "up to 2.5x inference gains" from software since the Spark's launch — a vendor claim, and one that is plausible for prefill and implausible for dense decode, because decode is bandwidth-bound and software cannot add bandwidth.

FAQ

Is the DGX Spark faster than a Strix Halo mini PC?

On the workload both are actually bought for, yes, by roughly a third: published llama.cpp benchmarks put gpt-oss-120b at 60.6 tok/s on GB10 hardware versus ServeTheHome's ~45 tok/s on AMD's Ryzen AI Halo Developer System. The gap tracks the bandwidth spec (273 versus 256 GB/s) plus a more mature runtime. It is not a large enough gap to justify a $4,699 Founders Edition against a $1,999-class Strix Halo box; it is arguably enough to justify the $3,099.99 GX10 clone if you also want CUDA.

Is a Mac Studio better than a DGX Spark for local LLMs?

For dense models, decisively — Apple publishes 819 GB/s on the M3 Ultra against NVIDIA's 273 GB/s on GB10, and dense decode scales almost linearly with bandwidth. LMSYS measured dense Llama 3.1 70B at 2.7 tok/s on the Spark; llmcheck.net lists 15 tok/s for Llama 3.3 70B on a 128GB M5 Max at 614 GB/s. For mixture-of-experts models the picture narrows and the Spark's CUDA stack becomes a real amenity. The Mac's hard limitation is that it runs no CUDA software at all.

What is the cheapest way to get 128GB of unified memory?

A Strix Halo mini-PC. The GMKtec EVO-X2 lists at $1,999.99 for 128GB, which is the lowest price per gigabyte in this whole comparison — but every configuration showed unavailable when we checked in early August 2026, and the cheapest genuinely in-stock box was the Minisforum MS-S1 Max at $3,639. Set restock alerts if you can wait; anything at or under roughly $3,000 for 128GB is a good outcome in this market.

How much of a 128GB Strix Halo box can a model actually use?

About 96-110GB, but only after you change the BIOS. The default allocation to the integrated GPU is roughly 16-32GB, which is why a 70B model will refuse to load on a 128GB machine straight out of the box. Allocate 96GB for LLM-only use and leave 32GB for the system; 110GB is the maximum practical allocation if you need a 70B at BF16 with long context.

Does the DGX Spark run dense 70B models at usable speed?

No. LMSYS measured 2.7 tok/s on Llama 3.1 70B at FP8 with SGLang at batch 1, and the arithmetic explains why: 70GB of weights against 273 GB/s gives a theoretical ceiling of about 3.9 tok/s. A Q4 70B (roughly 40GB) raises the ceiling to about 6.8 tok/s, which is still below comfortable reading pace. No software update changes this — if dense 70B chat is the goal, buy the Mac.

Should I wait for the next generation?

There is one dated thing worth knowing about: AMD announced Gorgon Halo (the Ryzen AI Max PRO 400 series) around 20 May 2026, with up to 192GB of LPDDR5X-8533 and up to 160GB allocatable as GPU memory at roughly 273 GB/s — about 7% more bandwidth than today's Strix Halo, with OEM systems expected in Q3 2026. If 192GB is what you need, that is a real reason to wait a quarter. If you are buying for bandwidth rather than capacity, note that the increment is 7%, not a category change.

Sources

• NVIDIA DGX Spark official page — GB10 Grace Blackwell, 128GB LPDDR5X coherent unified memory, 273 GB/s, ConnectX-7 networking
• AMD Ryzen AI Max product page — Ryzen AI Max+ 395: 16 Zen 5 cores, 40-CU Radeon 8060S, LPDDR5X-8000 on a 256-bit bus
• Apple Mac Studio technical specifications — M3 Ultra 819 GB/s, 96GB configuration; M4 Max 410 GB/s in the base Mac Studio
• llama.cpp DGX Spark benchmark thread — ggerganov (maintainer), Oct 2025: gpt-oss-120b 60.6 tok/s, Qwen3-Coder-30B 44.3 tok/s, prefill figures
• LMSYS DGX Spark benchmarks — SGLang and Ollama suite: dense Llama 3.1 8B and 70B at FP8, batch scaling
• ServeTheHome, AMD Ryzen AI Halo Developer System review (July 2026) — gpt-oss-120b ~45 tok/s, Qwen3.6-35B 62 tok/s, dense Qwen3.6-27B ~14 tok/s
• ServeTheHome DGX Spark review — teardown, 200GbE RDMA two-node clustering, launch-day performance caveats
• llmcheck.net Apple Silicon leaderboard — M5 Max 128GB: Llama 3.3 70B 15 tok/s (MLX), DeepSeek R1 70B 11 tok/s (Ollama), Qwen 32B-A3B 78-82 tok/s
• Apple MLX, set_wired_limit documentation — the iogpu.wired_limit_mb sysctl and the "strictly less than total memory" warning
• openai/gpt-oss-120b model card — 117B total parameters, 5.1B active per token, MXFP4
• Prices: vendor stores checked in the first week of August 2026 (see our Strix Halo mini-PC roundup); GB10 and Founders Edition pricing from our mid-July 2026 survey in GPU prices and the memory shortage