Thinking About a $5,900 RTX Spark Laptop? Test It on a DGX Spark First
TL;DR: The 128 GB RTX Spark laptops cost more than a DGX Spark: $5,899.99 for the top Surface Laptop Ultra and $6,229 for the HP OmniBook Ultra 16 The Register tested, against $4,699 for the desktop box. Both run the same chip, so a rented DGX Spark at $0.85/hour can tell you which memory size you actually need, and how fast your models will run, before you pay for the laptop. It cannot tell you about Windows, battery, or how the laptop handles heat. For those, read the reviews.
Last updated: October 2026. RTX Spark laptops started shipping this month and prices are moving. Check the manufacturer's store before you buy.
The first RTX Spark laptop reviews landed in mid-October 2026, and they settled one question that the spec sheets couldn't: price. Before launch, analysts expected RTX Spark machines to undercut the DGX Spark. The entry configurations do. A Surface Laptop Ultra starts at $2,599.99. But that model has 24 GB of memory and the cut-down 18-core chip. The configuration that matches a DGX Spark, 20 cores and 128 GB, costs $5,899.99. HP's equivalent was $6,229 as tested.
That is a $3,300 spread inside a single laptop model, and the only thing most buyers need to decide it is one number: how much memory do my models need? You can measure that number before you buy, on the same silicon, for the price of a few hours of rental.
The Short Version
- The RTX Spark's N1X and the DGX Spark's GB10 are the same chip. A model that loads and runs on a DGX Spark will load and run at a very similar speed on a 20-core N1X laptop.
- The expensive decision is the memory size. Configurations run from 24 GB to 128 GB. Measure what your model, at your context length, needs. Then buy that tier and not the one above it.
- Rent for the AI question, read reviews for the laptop question. A day of DGX Spark rental is $20.40 at $0.85/hour. The $5,900 laptop is about 6,900 rental hours.
Same Chip, Same Numbers
NVIDIA CEO Jensen Huang confirmed on record that the GB10 in the DGX Spark and the N1/N1X in RTX Spark systems are identical silicon: a 20-core Grace CPU, a Blackwell GPU with 6,144 CUDA cores, and up to 128 GB of unified LPDDR5X. The full comparison is in DGX Spark vs RTX Spark.
The Register benchmarked an RTX Spark laptop against a DGX Spark on the same model and engine (llama.cpp, single request), and the two land close together:
| Qwen3.6-35B-A3B | Generation speed | Time to first token |
|---|---|---|
| HP OmniBook Ultra 16 (RTX Spark, N1X) | 67.6 tok/s | 208 ms |
| DGX Spark (GB10) | 63.8 tok/s | 192 ms |
Source: The Register, October 2026.
The laptop is about 6% faster at generating text, which matches its slightly faster memory (about 300 GB/s against 273 GB/s). The DGX Spark is slightly faster to the first token, which The Register puts down to its larger chassis and better cooling. Neither gap changes which models feel usable. If you've wondered why generation speed tracks memory bandwidth so closely, prefill vs decode explains it.
So a measurement taken on a DGX Spark is a good stand-in for a 20-core RTX Spark laptop on short runs. For the 18-core N1X, which has 5,120 CUDA cores and tops out at 64 GB, expect somewhat slower prompt processing.
Why the Memory Size Is the $3,300 Question
On a DGX Spark, Linux hands memory to the GPU as it's needed. On Windows it doesn't work that way. The RTX Spark splits its memory into a CPU share and a GPU share. Reviewers found the 128 GB laptops ship with a 64/64 split, which leaves the GPU 62.5 GB. You can raise the GPU share in the NVIDIA app, up to 96 GB, but never past the memory you paid for. llama.cpp can spill a model into the CPU share, and because the memory is shared the penalty is smaller than on a discrete GPU, but it is still a penalty.
That gives you four practical bands for a model plus its context (the KV cache):
| What your workload needs on the GPU | What it means for the laptop |
|---|---|
| Up to ≈20 GB | You don't need RTX Spark's memory. A 24 GB discrete GPU laptop generates about 2.3× faster on models that fit. |
| ≈20–60 GB | Fits the default GPU share of a 128 GB laptop. A 64 GB configuration may also work if you leave enough for Windows and your apps. |
| ≈60–96 GB | 128 GB laptop only, with the GPU share raised in the NVIDIA app. |
| Over 96 GB | No RTX Spark laptop will hold it. A DGX Spark can, and two clustered DGX Sparks pool 256 GB. |
The bands are wide on purpose. The same model can land in two different rows depending on how it's quantized and how much context you run. A mixture-of-experts model like Qwen3.6-35B-A3B is about 20 GB at 4-bit, and a 122B model at 4-bit is about 70 GB. Add a long context window and either one moves up a row. Neither number tells you where your workload lands. You have to load it and look. For a broader guide to model sizes, see what fits in 128 GB.
How to Test It: One Afternoon on a Rented DGX Spark
Rent a DGX Spark through spark.enverge.ai. You get SSH and Docker on real GB10 hardware. Then:
- Pick the models you'd actually run on the laptop, at the quantization you'd use and the longest context you'd realistically need. A coding agent working through a repository needs far more context than a chat window.
- Load each one with llama.cpp. It is the engine The Register used, and it runs natively on RTX Spark Windows machines too. When a model loads, llama.cpp prints the size of the model buffer, the KV cache buffer, and the compute buffer. Add the three: that is roughly what the GPU share on the laptop needs to hold.
- Measure speed with
llama-benchat your prompt length. The generation number carries over to a 20-core laptop within a few percent on short runs. - Run your real workflow for a few hours. Point your coding agent, assistant, or data pipeline at the local model and use it the way you would on the laptop. This is where you find out whether 30 tokens a second is pleasant or painful for your work.
- Read your number against the table above. Buy the smallest tier that holds it, with headroom for Windows and the apps you keep open.
If you want to try several models across a few days, reserved instances cover 1–5 day windows. Five full days at $0.85/hour is $102, about 1.7% of the laptop's price.
What a DGX Spark Can't Tell You About the Laptop
Same chip, different machine. A rented DGX Spark answers the model-and-memory question, and only that question. Read the reviews for the rest.
- Windows on Arm. The RTX Spark runs Windows 11 on Arm, with x86 apps running through emulation. The Register notes PyTorch support is better on Linux than on Windows and hit driver problems. Check that your tools run.
- Heat and sustained load. The laptop N1X runs at 45–80 W, and the same chip in the DGX Spark is rated at 140 W. The Register saw GPU temperatures above 90°C and lockups after about an hour of sustained agent work. StorageReview's Surface lost power three times under a heavy all-core load, and a replacement unit did the same. Your DGX Spark numbers are a good guide for short bursts and an optimistic one for long jobs. If your work is hours of fine-tuning, a laptop is the wrong machine.
- Everyday laptop things. PCWorld found single-core performance weak. Battery life, screen, and keyboard are reviewers' territory.
What the Test Usually Tells You
- Your models fit in about 20 GB. You're paying for memory you won't use. Look at a cheaper RTX Spark configuration, or a discrete GPU laptop if raw speed matters more.
- Your models need 20–60 GB and you want them on the machine you carry. The 128 GB RTX Spark is the right purchase, and now you know it.
- Your models need more than 96 GB, or your work runs for hours at a time. No laptop fits. A desk-side DGX Spark, two of them clustered, or renting by the hour handles it. The buy vs rent breakdown covers that choice.
- You need the faster pipe more than the bigger tank. If your model fits in 96 GB and generation speed is the bottleneck, see DGX Spark vs RTX Pro 6000.
The point isn't that you shouldn't buy the laptop. For local AI on the Windows machine you live on, it may well be the right buy. The point is that a $5,900 decision with one unknown number in it is cheap to settle first. Start a DGX Spark instance, load your model, and read the number.
FAQ
Can a DGX Spark show how an RTX Spark laptop will run my model?
Yes, for the AI workload. The DGX Spark's GB10 and the RTX Spark's N1X are the same silicon, and in The Register's tests a DGX Spark ran Qwen3.6-35B-A3B at 63.8 tok/s against 67.6 tok/s on an RTX Spark laptop. It won't show Windows compatibility, battery life, or how the laptop holds up under hours of load.
How much memory does an RTX Spark laptop need for local LLMs?
As much as your model plus its context needs on the GPU, plus headroom for Windows. A 128 GB RTX Spark gives the GPU 62.5 GB by default and up to 96 GB if you change the split in the NVIDIA app. Load your model on a DGX Spark with llama.cpp and add the reported buffer sizes to find your number.
Is an RTX Spark laptop cheaper than a DGX Spark?
Only in the smaller configurations. RTX Spark laptops start around $2,600 with 24 GB, but the 128 GB configuration that matches a DGX Spark costs $5,899.99 (Surface Laptop Ultra) to $6,229 (HP OmniBook Ultra 16 as tested), against $4,699 for the DGX Spark.
Enverge provides cloud access to NVIDIA DGX Spark hardware for AI researchers, engineers, and startups. Use it for LLM inference, fine-tuning, benchmarking, and Blackwell-native development without buying the hardware upfront.