DGX Spark vs AMD Strix Halo: The $4,000 Local-AI Box Showdown

I published this comparison on 15 September. Seven days later two of its numbers are already wrong, and one of them is wrong in a way that matters. So this is a rewrite rather than a touch-up: new pricing, a much larger benchmark set, and a correction to the headline claim I made the first time.

The short version hasn't changed. These are the two 128GB unified-memory boxes an operator actually chooses between, and for most people the AMD machine is the better buy. What's changed is how wide the gap is, what you'll pay, and one architectural fact that matters more than either.

The two contenders, in plain terms

NVIDIA's DGX Spark is a GB10 Grace Blackwell superchip in a 150mm box that weighs 1.2kg. Twenty Arm cores, a Blackwell GPU, 128GB of coherent LPDDR5X, and — the part people miss — a ConnectX-7 NIC running at 200Gbps, which exists so you can lash two of them together.

Strix Halo is AMD's Ryzen AI Max+ 395: sixteen Zen 5 cores, a 40-compute-unit Radeon 8060S on RDNA 3.5, and the same 128GB of LPDDR5X across a 256-bit bus. It ships inside a dozen different machines from Framework, Bosgame, Minisforum, GMKtec, Beelink, Corsair, HP and ASUS, which is why its price varies by $1,800 depending on whose box you buy.

On paper the bandwidth difference is 7%. Hold onto that, because almost everything people say about these two machines being far apart in speed is not explained by 273 versus 256.

What you'll actually pay today

This is the section that dates fastest, so it carries a date: verified 22 September 2026.

Every one of those numbers is higher than it was three months ago, and the cause is not vendor greed. A DRAM contract-price shortage ran through 2026 and repriced the entire 128GB category. NVIDIA raised the Spark from $3,999 to $4,699 in February and said so plainly, citing "industry wide memory supply constraints." Framework's 128GB desktop went $1,999 at launch, $2,459 in January, and $3,449 today — and that last figure already reflects a price cut in September, after Framework secured cheaper inventory. Beelink's GTR9 Pro opened pre-orders at $1,985 and now sells for $4,349. Corsair's workstation roughly tripled.

The practical warning: almost every comparison of these machines still ranking in search quotes prices from the first half of the year. If you read that a 128GB Strix Halo box costs $1,999, that page has not been updated since June.

The benchmark that matters, and the one I got wrong

There are two numbers in local inference and most reviews only give you one.

Prefill is how fast the machine reads your prompt before it says anything. It is compute-bound. Decode is how fast tokens come out afterwards. It is bandwidth-bound. A single "tokens per second" figure hides which one you're looking at, and these two machines differ far more on one than the other.

Here is gpt-oss 120B at MXFP4, the model that actually needs this much memory:

Two corrections to what I wrote the first time.

The prefill gap depends on the AMD backend, badly. The same silicon does 431 tok/s under Vulkan and 818 under a ROCm build — nearly double, on identical hardware, from a compiler flag. Against the Spark's llama.cpp figure that's a gap of 2.1x at the tuned end and 4.5x at the untuned end. I quoted the untuned end as though it were the number. It isn't. If you buy one of these and care about long prompts, which backend you build is worth an afternoon.

Decode falls as context fills, and my earlier figure was a deep-context one. The Spark does 52.9 to 60.6 tok/s on an empty context, 44.8 at 16k, and 38.8 at 32k. The 38.6 I quoted originally was effectively the 32k number sitting next to a shallow-context AMD number. Comparing like with like: at low context it's roughly 55 versus 34, and at long context both fall.

So the honest summary is: the Spark is meaningfully faster at reading prompts, somewhat faster at generating, and both gaps are smaller than the first version of this article implied.

The thing that decides this more than either machine

If you take one thing from this piece, take this one, because it swamps the hardware difference entirely.

A mixture-of-experts model activates only a fraction of its weights for each token. A dense model activates all of them. On a machine whose bottleneck is memory bandwidth — which describes both of these boxes — that difference is enormous.

Read the Framework row again. The mixture-of-experts model has more total parameters than the dense one and runs seven and a half times faster. The 7% bandwidth difference between these two machines is noise next to that.

Which means the buying question isn't really "Spark or Strix." It's "what am I going to run?" If the answer is gpt-oss 120B, Qwen3 30B-A3B, GLM 4.5 Air or any other MoE, both machines are pleasant and the cheaper one wins. If the answer is a dense 70B, both machines are unpleasant — four to five tokens a second, which is slower than you can read — and no amount of choosing between them fixes that. You need a different class of machine, or a different model.

The one caveat on that table: the DGX Spark row comes from a published run that reports its figures but states no build hash or test date, unlike the other two. I've left it in because the pattern is consistent across every machine anyone has measured, but treat that specific 8.3x as the least firm number here.

The memory you can actually use

Both machines say 128GB. Neither gives you 128GB.

On Windows, Strix Halo is hard-capped at 96GB of graphics-addressable memory — 75% of system RAM, and vendors state it plainly in their own product titles. On Linux the picture is much better: memory is mapped rather than partitioned, so you keep the BIOS carveout tiny and raise the GTT limit instead, which gets you to roughly 110–120GB. AMD documents this in the ROCm system-tuning guide.

That is a real 24GB swing on the same hardware, and it is the strongest argument for running Linux on these boxes that has nothing to do with preference. If you intend to run Windows, budget 96GB and plan around it.

The Spark gives you about 120GB of its 128GB coherent pool and doesn't ask you to configure anything, which is worth something.

What each one is actually for

The October elephant

NVIDIA's RTX Spark N1X was confirmed in early September: 20 Grace cores, 6,144 CUDA cores, up to 128GB of LPDDR5X, eight OEMs, launching in October. It is the obvious reason to wait.

Two things about it are still genuinely unknown, and plenty of sites will tell you otherwise.

There is no announced price. Not from NVIDIA, not from any OEM. Lenovo, the first partner with detailed products, said explicitly that pricing comes at launch. Figures around $2,000 are journalist estimates and a $2,899 "leak" traces to a non-primary source.

There is no published memory bandwidth. If you see 600 GB/s quoted — and you will, it's everywhere — that figure comes from an NVIDIA slide reading "Unified Memory With NVLink C2C, 600 GB/s GPU to CPU." That's the die-to-die interconnect between the CPU and GPU, not memory bandwidth. NVIDIA has published no GB/s figure for the N1X at all. The closely related GB10 in the DGX Spark runs at 273 GB/s, which is the order of magnitude to expect.

So "wait for the N1X" is reasonable advice with an asterisk: you're waiting for a machine whose price and memory bandwidth are both unpublished, in a market where every comparable product got more expensive this year.

Every machine mentioned here, with today's price and the full benchmark set, is in the local AI hardware table — re-verified monthly.