A Third 128GB Mini PC? NVIDIA’s RTX Spark Leaves Out the One Number That Decides Local LLM Speed

September 7, 2026

I run local LLMs on a mini PC with 128GB of unified memory, the GMKtec EVO-X2 (hereafter, the EVO-X2). In September 2026, a second machine with the same 128GB joined the field.

NVIDIA announced “RTX Spark" on 3 September, with N1X as the SoC behind the first wave, and says machines built on it will ship in October. Looking only at the “128GB" printed on the spec sheet, it looks the same as the EVO-X2. Running local LLMs at home, would the two actually run at the same speed?

This is a survey as of September 2026.

What I Looked Into

The RTX Spark does not exist as a physical unit yet. I cross-checked the following three things.

• NVIDIA’s announcement from 3 September, and the spec sheet the company publishes for the DGX Spark (primary sources)
• Articles reporting on N1X’s internal breakdown (the part NVIDIA has not disclosed)
• The relationship between memory bandwidth and generation speed I previously measured across four machines of my own (measured)

What NVIDIA’s RTX Spark (N1X) Announcement Says

What NVIDIA has made public comes down to these three points.

According to reporting, N1X splits into two variants: the higher one with a 20-core Grace CPU, a 6,144-core GPU, and 24–128GB of unified memory, aimed at laptops and mini PCs; the lower one with an 18-core CPU, a 5,120-core GPU, and 24–32GB, aimed at laptops. This two-way split does not appear in NVIDIA’s own announcement text.

Looking at this from the position of someone running local LLMs, there is one number missing: memory bandwidth.

The DGX Spark’s Spec Sheet Has That Number

NVIDIA already has a product for AI developers in a small form factor, the DGX Spark, and its spec sheet is public.

Lined up side by side, the CPU core count and the compute figure both come out the same. This suggests a closely related chip, but N1X’s own memory bandwidth does not appear anywhere in NVIDIA’s published materials.

What Decides Speed Is Not Capacity

Why care about bandwidth? I have a result from measuring four machines of my own.

When an AI writes text, it re-reads the model’s weights from memory every time it produces one token (a fragment of a word). The amount read stays roughly the same each time, so how much memory can be read per second shows up directly as writing speed. That readable amount is memory bandwidth.

tok/s is the number of tokens written out per second. A token is a small chunk of text; in Japanese, roughly one to two tokens make up one character. The rightmost column, “generation speed ÷ bandwidth," lands in a 0.139–0.169 range for all four machines — a discrete GPU, a mini PC’s integrated GPU, and a Mac all included. With the same model and the same quantisation, generation speed scales roughly with memory bandwidth (Ollama, qwen3:8b and qwen3:14b, measured June 2026).

Capacity, on the other hand, decides whether a model fits at all, not how fast it runs. The number “128GB" means a large model fits — it does not mean it runs fast.

If You Know the Bandwidth, How Far Can You Predict Speed?

Applying the relationship above, generation speed can be roughly estimated from bandwidth. Since N1X’s bandwidth is not disclosed, I set up a scenario: what if it matched the DGX Spark’s 273 GB/s?

If this estimate holds, the RTX Spark and the EVO-X2 would run at nearly the same speed for 8B-class models. The two have very different compute figures — 1 PFLOPS is a large number — but writing text is capped by how fast memory can be read, not by compute.

There is room for this estimate to miss. The step where input is read (prefill) is decided by the amount of computation, so higher compute has the advantage there. A different way of holding the model, such as an MoE architecture, or a different implementation, would also shift things away from this proportional relationship.

Three Families Now Offer 128GB

By capacity, Apple leads; by price and availability, Strix Halo; and if you can wait until October, RTX Spark. But because RTX Spark’s bandwidth is unknown, speed is not something that can be compared at this stage.

What This Article Does Not Confirm

• I have not touched an actual RTX Spark unit. Machines built on it arrive from October
• N1X’s memory bandwidth is not disclosed; the estimate here assumes it matches the DGX Spark. It is not a measured value
• The relationship between bandwidth and generation speed comes from running qwen3:8b and qwen3:14b under Ollama. A different model or a different runtime could change the slope of that relationship
• The speed of the input-reading step was not included in this estimate

In Summary: The Same 128GB, Decided by a Different Number

• NVIDIA announced RTX Spark (N1X) on 3 September: a 20-core Grace CPU, a 1 PFLOPS Blackwell GPU, and up to 128GB of unified memory. Machines arrive from October
• NVIDIA’s announcement materials are missing memory bandwidth. The company’s own DGX Spark is listed at 273 GB/s, and the CPU core count and compute figure line up with N1X’s
• Across my own measurements on four machines, generation speed ÷ memory bandwidth held nearly constant at 0.139–0.169. Generation speed is roughly decided by bandwidth
• At 273 GB/s, an 8B-class model would generate at roughly 42 tok/s by this estimate — close to the EVO-X2’s measured 39.7 tok/s, and not a large difference
• The number “128GB" decides whether a model fits, not how fast it runs

Capacity and compute are what get printed large on the spec sheet, while bandwidth either sits small near the bottom of the table or is left off altogether. For someone bringing a model home to run it themselves, having the one number you most want to see stay unknown until the last minute makes for an unsettled kind of hardware shopping.

For a look at how far Apple’s “4.3x AI performance" actually reaches into local LLMs, here is a related article.

Apple Says 4.3x AI Performance. Which Half of a Local LLM Does That Touch?

Last time I lined the Mac Studio up against other machines and guessed that capacity is what mat ...

 https://kimono-oyaji.com/mac-neural-mlx-en/

Sources

• NVIDIA official blog (3 September 2026; content of the RTX Spark announcement)
  https://blogs.nvidia.com/blog/local-ai-ifa-next-gen-agents-nv-pair-rtx-spark/
• NVIDIA DGX Spark product page (source for the 273 GB/s memory bandwidth and other specs)
  https://www.nvidia.com/en-us/products/workstations/dgx-spark/
• Hermitage Akihabara (4 September 2026; source for the report that N1X splits into two models)
  https://www.gdm.or.jp/pressrelease/2026/0904/650789

Hardware Featured in This Article

This is the 128GB mini PC I own, used here as the baseline for comparison.

As an Amazon Associate we earn from qualifying purchases.