A compact desktop AI computer sits beside an exploded view illustrating its chip and cooling system.
The XDA Developers review of the MSI EdgeMesa N AI+ makes a bold claim. It says the mini PC form factor lets Nvidia's RTX Spark chip run at "full whack." The claim holds up, with caveats. The short version is that a compact desktop can run this Windows-on-Arm chip at a higher power limit than a laptop can. For local LLM work, that headroom matters most when many requests hit the box at once. It matters much less when one person chats with a small model.

A compact desktop AI computer sits beside an exploded view illustrating its chip and cooling system. What launched, and when​

Hothardware notes that the mini PC has no battery and more room for heatsinks, fans and cooling. Its RTX Spark therefore gets a tweaked power profile with a 140 watt TDP. Techtroduce, reporting on ComputerBase's testing, says the EdgeMesa drew about 105W of SoC power in Gears of War: E-Day at 1440p Ultra. An Asus ProArt P16 laptop with the same processor drew 63W.

Per Nvidia's launch post, MSI calls the thermal design the AI Pulse Cooling System. It is a vapor chamber with two fans that MSI rates to sustain 140W.

That 140W is a vendor rating for MSI's chassis. It is not a promise for every RTX Spark mini PC. XDA's review notes that systems from Asus, Dell and MSI are available now. Acer, Gigabyte, HP and Microsoft are listed as coming soon.

The hardware​

The platform specs are shared across the RTX Spark desktops:

  • A 20-core Nvidia Grace Arm CPU.
  • A Blackwell RTX GPU with 6,144 CUDA cores.
  • Up to 128GB of LPDDR5X unified memory.
  • Windows 11 on Arm.

The MSI EdgeMesa N AI+ reviewed by XDA had 128GB of RAM and a 4TB Gen5 SSD. XDA lists these details for the unit:

  • It measures 152 x 147 x 53 mm and weighs 1.32 kg.
  • It has four 20 Gbps USB-C ports, HDMI 2.1 and 10GbE.
  • It has Wi-Fi 7 and Bluetooth 5.4.

StorageReview's teardown adds some practical points. One of the four USB-C ports is the power input. The system also has a single M.2 slot that takes only the short 2242 format, so the capacity you buy is the capacity you keep.

Does the full TDP actually pay off?​

Independent testing says yes, but modestly, and only under some loads.

  • Sustained CPU work: StorageReview measured 4 to 11 percent more multi-core throughput than the fastest N1X laptops. It also measured 18 to 25 percent more output when eight requests hit a 27B model.
  • Single-user work: StorageReview says single-request results landed within about 6 percent across every system it compared. For one person chatting with a small model, the mini PC and the laptop feel the same.
  • Scaling with power: Techtroduce, citing ComputerBase, found that higher power did not translate into proportional gains. GPU clocks were about 8% higher in 3DMark Steel Nomad.
  • A different OEM: Windows Central's ASUS ProArt GR1X review is a separate system. It found the desktop chip's TDP tops out at 140W against 80W for the laptop version. It also found gains of roughly 5% in multi-core CPU tests and up to 13.3% in some 3DMark tests.

XDA reaches a similar conclusion. Performance hasn't changed much from the laptop versions, though GPU-bound tasks got a small uplift. The reviewer also reports that the unit was very quiet and that temperatures in HWiNFO64 stayed low.

StorageReview adds a caveat. The EdgeMesa averaged 88.4C in its sustained loop, but its peaks reached 107.3C. The ASUS unit ran cooler on average. Quiet and cool-on-average does not mean the chip never gets hot.

The "128GB of VRAM" question​

XDA's own pros and cons list says the memory is "not truly unified." That matters for anyone sizing a model.

StorageReview's unit, at its default allocation, showed 62.5GB as dedicated GPU memory and 47.9GB as shared. It describes it as one physical pool that Windows splits into a dedicated GPU portion and system memory. The GPU can spill into system memory when it needs more.

A slider in the Nvidia app changes the split. StorageReview says Nvidia advises keeping the default unless a workload needs more. A bigger GPU share leaves less for Windows, and Nvidia recommends leaving at least 16GB for the CPU. For image and video generation models, Nvidia suggests raising the allocation. Agents and LLMs, it says, lose little performance when they spill into system memory.

In practice, 128GB gives you more headroom, not 128GB of dedicated VRAM. StorageReview says the 128GB unit ran a 35B coding-agent model and Qwen-Image 20B at Q8. Its Windows suite did not run those on the 64GB systems.

Local LLM results​

The numbers below come from different reviewers using different tests. Don't mix them.

  • StorageReview, llama.cpp CUDA: One request on the EdgeMesa gave 58.1 tokens per second on a 3B model and 13.0 on Qwen3.8 27B. With eight requests on that 27B model it produced 39.8 tokens per second, against 33.8 and 31.9 on the two laptops.
  • Where it's weaker: StorageReview found the EdgeMesa slightly behind the laptops in FLUX.2 klein image generation. It suspects the cause is a newer suite version rather than the hardware.
  • Untested claim: MSI advertises support for 120B-parameter models. StorageReview's Windows suite doesn't include models that large, so it did not verify the claim.
  • How-To Geek vs Strix Halo: Qwen3.5 35B-A3B ran at 65.7 tok/s on the Spark, against 62.3 tok/s on the Strix Halo. The reviewer pairs that with a warning that Windows on Arm is the system's biggest limitation.

The pattern is consistent. Text generation is a close race with AMD's Strix Halo. The Spark pulls ahead when CUDA-friendly work such as image generation comes into play.

Price and trade-offs​

Pricing is where reports differ, so check current listings before you buy.

  • XDA lists the 128GB/4TB MSI at $8,000.
  • StorageReview and How-To Geek both cite $7,499 for that configuration. StorageReview attributes the figure to Nvidia and adds that RTX Spark desktops start at $4,399.
  • XDA says Microsoft's Surface RTX Spark Dev Box is $6,000 with 128GB and a 2TB SSD, but it is not yet available.

XDA's complaints are practical:

  • All four USB-C ports are 20 Gbps, not 40 Gbps USB4.
  • There's no USB-A port and no front I/O.
  • There's no ConnectX-7 networking, so the 10GbE port suits a single box rather than a cluster.
  • Windows is not a pleasant OS to run headless and remote. The reviewer says you might be better off with a DGX Spark in that case.
  • Windows on Arm still has rough edges. StorageReview reports that an x64 Cinebench build crashed under emulation on all four of its test systems.

Who should buy one​

This is a niche purchase. The case for it is strongest if you:

  1. Develop or run local AI on Windows with CUDA and WSL.
  2. Want a dedicated always-on box for overnight agent tasks.
  3. Need large memory capacity more than raw benchmark wins.

Nvidia itself notes that native CUDA support and Linux tooling through WSL are the foundation for building local workflows. If you want clustering, a Linux-first setup or a lower price, the alternatives look better. The alternatives are the DGX Spark-based boxes, a 64GB RTX Spark machine, or an RTX Spark laptop. XDA scored the EdgeMesa 8/10. That is a fair score for a platform that is capable and expensive, with a few Arm-shaped rough edges.

 

References

  1. RTX Spark mini PCs let you use the full TDP of the chip, and that's perfect for local LLM use XDA 2026-11-20T01:01:20+00:00
  2. MSI EdgeMesa N AI+ Review: RTX Spark N1X and 128GB in a 1.18-Liter Mini PC - StorageReview.com storagereview.com
  3. MSI EdgeMesa N AI+ Review: NVIDIA RTX Spark Power In A Pint-Sized PC hothardware.com