Yuval Avidani
Author
Turns out the little machine everyone's calling a "personal supercomputer" isn't a laptop at all. NVIDIA's DGX Spark is a box that sits on your desk, palm-sized, weighing about a kilo and a quarter, with an external 240-watt power supply. So if you've seen someone waving it around like it's some beefy laptop, that's mistake number one we're clearing up right here.
Let's break this down from the top. There are two boxes everyone's talking about in 2026: NVIDIA's DGX Spark, and AMD's Ryzen AI Max+ 395, codenamed Strix Halo. Both are aimed at exactly the same thing: running big AI models locally, at home, no cloud needed. And the question everyone's asking is simple: who runs a local LLM better?
The unsexy but correct answer is that in text generation speed, the two boxes are almost tied. And that's exactly the part marketing doesn't want you to know.
What is "unified memory" anyway, and why is it the whole story
Before we compare, we need one concept clear: unified memory. On a regular computer you've got RAM on one side and a graphics card with its own memory on the other, and the two have to shuttle data back and forth. In these boxes, the memory is shared by everyone: the CPU, the GPU, and the chip running the AI all reach into the same big pool of 128GB.
Why does this matter to us? Because a big model, say around 70 billion parameters, simply doesn't fit into any regular consumer graphics card on the market. That 128GB of unified memory lets you hold a whole model like that in memory at once. The DGX Spark comes with 128GB LPDDR5X at roughly 273 gigabytes per second of bandwidth. The Strix Halo also goes up to 128GB, at roughly 256 gigabytes per second, which is about 7 percent below NVIDIA.
And here's where the big mental trap kicks in: people think a 7 percent bandwidth gap is nothing, and that NVIDIA must be crushing it across the board. Well, no.
Why text generation is nearly a tie
Let's understand what happens when a model "writes" you an answer. That stage is called decode — generating tokens one after another. The bottleneck here isn't how many calculations the machine can do, it's how fast it can pull the model's weights out of memory. In other words: text generation is bandwidth-bound, not compute-bound.
Think of it like a waiter running back and forth to the kitchen. Doesn't matter how many hands he has — what matters is how fast he runs down the hallway. And if both boxes have a hallway of nearly the same width, they'll both serve food at roughly the same pace.
In an independent test on the 120-billion-parameter gpt-oss model, the Ryzen AI Max+ 395 hit 34.13 tokens per second versus 38.55 for the DGX Spark. An edge of about 13 percent for NVIDIA — not a generational gap. And if that's not enough, The Register found that with the Vulkan engine inside llama.cpp, AMD actually took a small lead in token generation. So anyone imagining that NVIDIA "wipes the floor" with the competitor in day-to-day local chat is simply wrong.
Where NVIDIA truly wins
Now, to be fair, there are two places where the Spark does open up a real gap.
The first is called prompt processing, or by another name, time-to-first-token: how long it takes the model to read what you wrote and start answering. This stage, called prefill, is the opposite of the previous one: it's compute-bound, not bandwidth-bound. And here NVIDIA's Blackwell GPU, together with CUDA, is simply much stronger. On a 256-token prompt the Spark is about 2 to 3 times faster, and in some tests up to 5 times. And the longer the context, the bigger the gap gets.
Why does this matter to us? Because if you're feeding the model long documents, or building an agent that reads a lot of text before it answers, that initial wait is noticeable. That's where the Spark gives you a different experience entirely.
The second is image generation. In a FLUX.1 Dev test, the DGX Spark, at roughly 125 TFLOPS in BF16, was 2.5 times faster than the Strix Halo, which sits around 46 TFLOPS. Again, a task that's purely compute-bound, and there Blackwell rules.
The price: double, and that's not a footnote
Now to the part that decides most real-world purchase decisions. The DGX Spark was announced on March 19, 2025 at GTC, originally under the name Project DIGITS, and went on sale on October 15, 2025 at an official price of $3999. In practice on the street it's more: about $4399 on Newegg, and about $4699 on NVIDIA's Marketplace bundled with a course.
On the other side, the Framework Desktop with 128GB, one of the leading Strix Halo machines, sits at about $1999 — roughly half the Spark's price. The HP Z2 Mini G1a, at about $2949, is also significantly cheaper. And beyond price there's a fundamental difference: the Strix Halo is, at the end of the day, a perfectly normal computer — you can work on it day to day. The Spark is a dedicated AI device.
Bottom line: don't fall for the "supercomputer" marketing
In my view, this whole story is really a budget decision dressed up as a tech monster. If all you want is to run local chat and talk to big models, AMD gives you about 90 percent of the experience at half the price, and it doubles as a regular computer too. But if you're building agents, doing fine-tuning, or generating images and video through CUDA and Blackwell, the Spark is an AI lab in a box that might just justify the extra spend.
And that's exactly what The Register said: for local LLM inference, they actually recommend Strix Halo systems — because of the price, because the bandwidth is roughly equal, and because they double as a real computer. The Spark, on the other hand, is a better fit for prototyping agents, fine-tuning, and image generation.
Bottom line: don't fall for the "personal supercomputer" marketing. Ask yourself first what you're actually going to do with the box, and only then look at the price. It's worth remembering that all the numbers here come from specific tests on specific models, and on a different model or with a different engine the picture might look a bit different. This isn't a buying recommendation — it's just a sober perspective on what these machines actually give you.
So now that you understand the speed is nearly identical, and the real gap is in price, the only question left is: are you actually building agents and training models, or do you just want to tell people you've got a supercomputer on your desk?