日本語

Showcase · Oct 16, 2026

Reading Artificial Analysis's First RTX Spark Laptop Results, Posted Just Before Launch

At dawn on a desk, a laptop screen lets streams of agent work flow out as long bands of light

On launch morning, NVIDIA's Japanese account announced that RTX Spark was on sale from that day (in Japanese) 1.

A few hours earlier, the benchmark lab Artificial Analysis had posted its first results for an RTX Spark laptop on X. On Qwen3.8 27B, it says, the laptop finished a recorded local agent workload in 18.7 minutes 2.

What was measured

The test is the lab's AA-AgentPerf-Local. It is an open-source tool for measuring how fast a laptop or workstation serves an AI agent, and it replays recorded agent sessions against a model server running on the machine 3.

The test machine was a prototype 15-inch Surface Laptop Ultra 4. According to Artificial Analysis, it was plugged in, with Windows set to high-performance mode, using NVIDIA's published llama.cpp recipes where available 4. So this is a conditional measurement on a prototype. It does not mean every shipping unit will land on the same result.

If you want to look up the retail version, start here.

Reading the results

The default workload on Qwen3.8 27B is 8 recorded agent tasks and 168 model turns 2. Artificial Analysis says the RTX Spark laptop finished in 18.7 minutes, close to DGX Spark at 17.3 minutes 2. Its comparison: 2.0x faster than the MacBook Pro (M5 Pro, 64 GB, 37.6 minutes) and 1.8x faster than the AMD Ryzen AI Halo (34.5 minutes) 2.

The fastest machine was the GeForce RTX 5090 desktop, at 4.9 minutes 5.

A second model: on Qwen3.6 35B A3B, the lab says the laptop takes 7.6 minutes, close to DGX Spark at 7.3 minutes and 1.6x faster than the MacBook Pro at 11.9 minutes 6.

The last one is Ling 3.0 Flash, which the lab calls the largest current model it tests 7. The laptop takes 15.3 minutes there, close to DGX Spark at 14.9 minutes and 1.6x faster than the Ryzen AI Halo at 25.1 minutes 7. And the lab writes that the model's 77 GB of weights do not fit on the MacBook Pro or on the RTX 5090 7.

That is the reading that matters. On Qwen3.8 27B, the RTX 5090 desktop was quickest 2 5. On Ling 3.0 Flash, its 77 GB of weights do not fit on the RTX 5090, while the RTX Spark laptop, DGX Spark and the Ryzen AI Halo, with their large memory pools, finish the run 7. The dividing line is not only speed but whether the model fits. RTX 5090 was not slow.

Memory and power: NVIDIA's documents

On the memory side, NVIDIA's documentation gives the details. According to NVIDIA, the memory is shared by the CPU and the GPU, on a 256-bit interface with 300 GB/s of bandwidth 8. Its product page lists laptop power (TDP) at 45 to 80 W 9.

Artificial Analysis sees RTX Spark's N1X as essentially the GB10 from DGX Spark (20-core Arm CPU, 6,144-core Blackwell GPU), with up to 128 GB of memory at 300 GB/s 10. That is the lab's reading of the chip, not a claim that the chips perform identically.

Big model weights also need somewhere to live on disk.

If you want to start at your own desk

To try something small at home, one way in is Llama Windows from Hugging Face, a free and open-source app that puts llama.cpp in your Windows 11 tray 11.

According to Hugging Face, it has native support for Windows on Arm 12 and for RTX Spark 13. They say they worked with NVIDIA to have it ready on RTX Spark's first day 14. It also includes a local server with an OpenAI-compatible API, so your own apps and agents can call the same models 15.

Still choosing the machine to install it on? You can search for RTX Spark laptops here.

PC Watch also ran a headline (in Japanese) saying a llama.cpp Windows app has appeared with day-one support for RTX Spark 16.

Before you buy

PC Watch ran a headline (in Japanese) that opens with "even at [the yen amount shown in the post below], sold out immediately" 17.

From Japan, a PC Watch headline (in Japanese) reports a 16-inch HP Japan laptop with RTX Spark, 15.7 mm thin 18.

Benchmark software is moving too. A 4Gamer headline (in Japanese) reports 3DMark 2.32.8919, released with stronger Windows on Arm support in light of RTX Spark's arrival 19.

Here are some machines. The first card shares a product name with the machine Artificial Analysis tested, but the tested unit was a prototype. Check that the software you want to run supports the machine before you choose.

Preorder at Microsoft StoreAdSearch on Amazon
AdSearch on Amazon

For comparison, here is DGX Spark.

Official product pageAdSearch on Amazon

Read next

Specs second, measurements first. For more on running AI at home, continue with What People Run on Their Own Machines.

Sources

  1. Post by @NVIDIAGeForceJPWindows PCの新たな幕開け ― RTX Spark が本日より発売開始。
  2. Post by @ArtificialAnlysOn Qwen3.8 27B the RTX Spark laptop finishes our default workload (8 recorded agent tasks, 168 model turns) in 18.7 minutes, close to DGX Spark (17.3), 2.0x faster than the MacBook Pro (M5 Pro, 64 GB, 37.6) and 1.8x faster than the AMD Ryzen AI Halo (34.5)
  3. AA-AgentPerf-Local: First AA-AgentPerf-Local results for an NVIDIA RTX Spark N1X laptop | Artificial AnalysisAA-AgentPerf-Local is our open-source tool for measuring how fast a laptop or workstation serves an AI agent. It replays recorded agent sessions against a model server running on the machine
  4. Post by @ArtificialAnlysWe tested a prototype 15-inch Surface Laptop Ultra (128 GB, $5,900), plugged in with Windows set to high-performance mode, using NVIDIA's published llama.cpp recipes where available
  5. Post by @ArtificialAnlysThe GeForce RTX 5090 desktop is fastest, at 4.9 minutes
  6. Post by @ArtificialAnlysOn Qwen3.6 35B A3B it takes 7.6 minutes, close to DGX Spark (7.3), and 1.6x faster than the MacBook Pro (11.9)
  7. Post by @ArtificialAnlysOn Ling 3.0 Flash, the largest current model we test, it takes 15.3 minutes, close to DGX Spark (14.9) and 1.6x faster than the Ryzen AI Halo (25.1). Its 77 GB of weights do not fit on the MacBook Pro or the RTX 5090
  8. NVIDIA RTX Spark: System Overview — NVIDIA Windows on Arm Porting GuideMemory: Up to 128 GB of LPDDR5, shared by the CPU and iGPU. The interface is 256 bits wide, with a bandwidth of 300 GB/s.
  9. NVIDIA RTX Spark: Slim Laptops & Small Desktops | NVIDIA RTX SparkPower (TDP) 45 - 80 W 45 - 80 W
  10. Post by @ArtificialAnlysRTX Spark's N1X chip is essentially the GB10 from DGX Spark (20-core Arm CPU, 6,144-core Blackwell GPU), with up to 128 GB of memory at 300 GB/s
  11. Llama Windows: Introducing Llama Windows: llama.cpp in your Windows trayToday we're announcing Llama Windows, a free and open source app that puts llama.cpp in your Windows 11 tray.
  12. Llama Windows: Introducing Llama Windows: llama.cpp in your Windows trayIt launches with native support for Windows on Arm and for the new NVIDIA RTX Spark.
  13. Llama Windows: Introducing Llama Windows: llama.cpp in your Windows trayLlama Windows launches with support for NVIDIA RTX Spark. It runs natively on Windows on Arm, and llama.cpp runs models on the RTX Spark GPU.
  14. Llama Windows: Introducing Llama Windows: llama.cpp in your Windows trayWe worked with NVIDIA to have Llama Windows ready for RTX Spark on day one.
  15. Llama Windows: Introducing Llama Windows: llama.cpp in your Windows trayYour own apps, scripts, and agents can use the same models the app loads, through its OpenAI-compatible API.
  16. Post by @pc_watchllama.cppのWindowsアプリ登場、RTX Sparkに初日対応
  17. Post by @pc_watch100万円でも即売り切れ。NVIDIA「RTX Spark」に触れて分かった、可能性と難しさ
  18. Post by @pc_watch日本HP、RTX Spark搭載で薄さ15.7mmの16型ノートPC
  19. Post by @4GamerNews「RTX Spark」の登場を踏まえてWindows on Arm対応を強化した「3DMark」バージョン2.32.8919リリース

Embedded posts belong to their authors. Amazon links are affiliate links: we may earn a commission. They open a search, not a specific listing.