What RTX Spark actually is
RTX Spark is not a normal gaming laptop with a discrete GPU. Nvidia's developer documentation describes an Arm-based system-on-chip for Windows laptops. The top configuration has a 20-core CPU. That figure comes from Nvidia's porting guide, not from the review search results. The CPU is built on Arm v9.2 and split into two 10-core clusters. Each cluster has five Cortex-X925 performance cores and five Cortex-A725 efficiency cores. The GPU is an integrated Blackwell part.
Memory is the key design choice. Nvidia's documentation says the CPU and GPU share up to 128 GB of fully coherent memory. It runs over a 256-bit interface at 300 GB/s. PCWorld says this is why you can't build an equivalent desktop yourself.
Not every laptop has the top configuration. Tom's Hardware's Surface Laptop Ultra unit had the top-end N1X chip with 20 CPU cores and 6,144 CUDA cores, but only 32GB of unified memory. ASUS lists ProArt P16 memory options of 24, 32, 64 and 128 GB. Nvidia says laptops start at $2,599 and that prices and configurations vary by region. That starting price doesn't necessarily buy a 128 GB model.
The benchmark that got everyone's attention
On the ProArt P16, Chacos says he could assign roughly 96 GB of the 128 GB pool to the GPU. He compared it with his desktop, which has 96 GB of system RAM alongside a GeForce RTX 5090 with 32 GB of GDDR7, for about 128 GB in total split across two separate pools. Both ran the same large research workload on OpenAI's 120-billion-parameter gpt-oss model.
| System | Total tokens | Time |
|---|---|---|
| ASUS ProArt P16 (RTX Spark, 128 GB) | 472,084 | 33 min 13 s |
| RTX 5090 desktop (128 GB RAM) | 472,084 | 52 min 51 s |
The table repeats the token count for the desktop. PCWorld states the count for the ProArt only, but the workload was the same on both. The laptop finished almost 20 minutes sooner, a gap of 19 minutes and 38 seconds.
Chacos is explicit that the laptop did not win because its GPU is faster. It isn't faster. The model was too big for the 5090's 32 GB, so the desktop had to spill into much slower system RAM. A workload that fits inside 32 GB of GDDR7 would let the desktop pull ahead. Read this as a result for one memory-hungry local AI job, not as "RTX Spark beats RTX 5090." It is also one reviewer's result on one workload.
Shared memory isn't the same as VRAM either. At 300 GB/s, the RTX Spark pool is far slower than the memory on a high-end discrete card. The advantage is capacity, not speed.
Gaming and everyday use
PCWorld's testers found the integrated Blackwell GPU roughly keeps pace with a discrete GeForce RTX 5070. DLSS upscaling and frame generation helped a lot in supported games. Mark Hachman saw Cyberpunk 2077 reach triple-digit frame rates with them on. But conventional GeForce laptops gave better gaming performance for much less money. Chacos found the extra memory made no difference to games.
Other reviewers agree on some points and differ on others:
- Tom's Hardware says a Windows power user won't get much out of RTX Spark beyond better software and gaming compatibility than earlier Arm laptops.
- The Shortcut notes that discrete GPUs have much faster video memory and that the Surface runs almost every game through Microsoft's Prism emulator.
- Notebookcheck calls the Surface Laptop Ultra a bit disappointing, especially given the high price.
For day-to-day work, Hachman found excellent multicore performance from the 20 Arm cores. Single-core performance was considerably weaker, and that matters for how responsive the machine feels. PCWorld's own Surface review lists slow single-core performance that may slow down the UI among its cons.
Battery results depend on the test. PCWorld got about 10 hours from the Surface in office-work testing. The Shortcut reported eight- and 10-hour days before needing a charge. Computerworld measured 14 hours 35 minutes in PCMark video playback. Video playback is a lighter load than office work, so these figures don't conflict.
Heavy AI work drains the battery fast. During Chacos's 33-minute run, the ProArt used 46 percent of a full charge, and its fans were very loud. Running a 120B model locally on a thin laptop works, but you'll want to stay near an outlet.
The Windows on Arm caveat
Windows on Arm has improved a lot. Microsoft's Prism translation layer handles much x86 software, and native Arm support keeps growing. But it isn't seamless yet. Hachman needed a beta version of Adobe Premiere because the shipping version didn't properly support the RTX Spark GPU. Nvidia's launch announcement says Photoshop and Premiere offer native RTX Spark support through beta releases at launch. Those details are consistent.
Check your software before you buy, especially if you depend on:
- specialized utilities
- drivers
- plug-ins
- legacy Windows applications
Nvidia's announcement also lists developer tooling for the platform, including a PyTorch-CUDA build for Windows on Arm and a native Python build. Those are Nvidia's claims, not PCWorld's findings.
Who should buy one
PCWorld's conclusion is blunt:
- AI developers and researchers who run models that overwhelm conventional GPU memory, or who spend heavily on cloud inference, should pay attention.
- Creative professionals doing a lot of local AI work may like it, provided their software works on the platform.
- Gamers should look elsewhere unless gaming is secondary to AI work.
- Everyone else can get a fast premium laptop for far less.
Chacos also warns that $7,000 buys a lot of cloud AI. That is a fair counterpoint. The case for local hardware rests on privacy, avoiding per-token costs and having a large model on hand offline. If you don't have a workload that fills that memory, the case is weak.
Nvidia markets up to 1 petaflop of FP4 AI performance and up to 128 GB of unified memory. Treat those as manufacturer specifications, not as results you'll see in every task.
The bottom line
RTX Spark is a first-generation platform with one clear strength. It can load models a thin laptop couldn't touch before. That strength comes with a high price, a heavy battery cost under load, and the usual early Arm compatibility issues. If you can name the workload that needs 96 GB of GPU-accessible memory, it may be exactly what you want. If you can't, a conventional laptop will serve you better and cost less.
Correction (November 4, 2026): The desktop's memory was originally described as 128 GB of system RAM. PCWorld's review states the desktop has 96 GB of system RAM plus the RTX 5090's 32 GB of GDDR7, about 128 GB in total across two separate pools. The text has been corrected.