On September 3, NVIDIA announced that Windows PCs built around its Arm-based processor, "RTX Spark," will go on sale in October 2026. The "fall" timeline given in the initial announcement on May 31 has now become more concrete, and the official specifications list both 18-core and 20-core configurations.
With up to 128GB of shared memory, the selling point is that even thin-and-light laptops can run CUDA and RTX-compatible software. However, AMD and Apple also offer products with comparable memory capacity. The reasons to choose RTX Spark come down to more than the amount of data it can handle: memory transfer speed and how much of your existing software actually runs matter too.
18-Core and 20-Core Versions, With Power Limits Depending on the Chassis
At the heart of RTX Spark is the NVIDIA N1X. NVIDIA's product page lists two laptop configurations: 20 CPU cores, 6,144 CUDA cores, and up to 128GB; and 18 CPU cores, 5,120 CUDA cores, and up to 64GB. Both have a power envelope of 45 to 80W. The desktop version, by contrast, has 20 CPU cores, 6,144 CUDA cores, and up to 128GB, with a 140W allocation. All configurations use LPDDR5X and Windows 11.
| Use | CPU cores / CUDA cores | Max shared memory / Power envelope |
|---|---|---|
| Laptop (high-end) | 20 / 6,144 | 128GB / 45–80W |
| Laptop | 18 / 5,120 | 64GB / 45–80W |
| Desktop | 20 / 6,144 | 128GB / 140W |
Even with the same 20 cores and 6,144 CUDA cores, laptop and desktop performance needs to be evaluated separately. A thin laptop's 45–80W and the desktop's 140W represent different design constraints. Identical core counts do not guarantee the same sustained speed in long inference or rendering workloads.
NVIDIA says it collaborated with MediaTek on CPU design and connects a Grace CPU and a Blackwell RTX GPU via NVLink-C2C. The CPU is Arm-based, and CUDA runs natively. The configuration pitches the size of the memory the GPU can use together with the ability to run software built for CUDA.
Lenovo's Actual Specs and Per-Model Launch Information
The concrete model to emerge from Lenovo's September 3 announcement is the 15-inch Yoga Pro 9n. It features up to 128GB of 256-bit, 9,400MT/s LPDDR5X, is as thin as 16.7mm and as light as 1.65kg, and has an 80W TDP. The design aims to fit a laptop with a large shared memory pool into a portable size.
Shared memory means the CPU and GPU use the same physical memory. That can help when running large models locally, but not all 128GB can be assigned to model weights. Windows, applications, and inference working space all draw on the same capacity. The capacity figure alone does not tell you how large a model you can run.
Lenovo also announced the 16-inch Yoga 9n 2-in-1, which features a 360-degree rotating chassis and pen support. However, pricing and availability for both the Yoga Pro 9n and Yoga 9n 2-in-1 are to be announced later. The announcement includes no Japan-specific pricing or release date.
Acer's compact desktop shown in NVIDIA's IFA announcement was also displayed only as a concept. The platform's October launch window does not mean every showcased model has a set release date. When narrowing down candidates, check each manufacturer's sales regions and configurations.
Same 128GB, Different Bandwidth and Software
Large shared memory is not unique to NVIDIA. AMD's Ryzen AI Max+ 395 offers 16 CPU cores and 32 threads, 40 Radeon 8060S GPU compute units, and up to 128GB of 256-bit LPDDR5X-8000. For the 128GB configuration of the Ryzen AI Halo, which uses this chip, memory bandwidth is listed at 256GB/s. The 395's chip specifications list support for Windows 11 and x86 Linux, with a default TDP of 55W and a configurable range of 45–120W. The AI Halo itself has a 120W TDP.
MacBook Pro models with Apple's M5 Max offer 18 CPU cores, up to 40 GPU cores, up to 128GB of shared memory, and up to 614GB/s. Qualcomm's Snapdragon X2 Elite Extreme, model X2E-94-100, lists Arm64-compatible 18 CPU cores, an Adreno X2-90 GPU, up to 128+GB of LPDDR5X (actual installed amounts vary by product), and 228GB/s. Its Hexagon NPU delivers 80 TOPS at INT8.
| Platform | Published CPU / GPU configuration | Max shared memory | Memory bandwidth | OS / main software conditions |
|---|---|---|---|---|
| NVIDIA RTX Spark (Yoga Pro 9n top spec) | Up to 20 CPU cores / up to 6,144 CUDA cores | 128GB | 300.8GB/s (theoretical) | Windows 11, CUDA/RTX |
| AMD Ryzen AI Max+ 395 | 16 CPU cores, 32 threads / Radeon 8060S, 40 compute units | 128GB | 256GB/s (AMD figure) | Windows 11, x86 Linux |
| Apple M5 Max | 18 CPU cores / up to 40 GPU cores | 128GB | Up to 614GB/s (Apple figure) | macOS, MacBook Pro |
| Snapdragon X2 Elite Extreme X2E-94-100 | 18 CPU cores / Adreno X2-90 | Up to 128+GB | 228GB/s (Qualcomm figure) | Windows on Arm, NPU 80 TOPS at INT8 |
The theoretical memory bandwidth calculated from the Yoga Pro 9n's published specifications is 300.8GB/s. The formula is: 9,400 MT/s × 256 bit ÷ 8 ÷ 1,000 = 300.8 GB/s. This value is derived from the configuration Lenovo disclosed for that model; it is neither a measured speed nor a guaranteed figure common to all RTX Spark products.
The table lines up each company's published specifications as confirmed on September 4 and is not a same-conditions benchmark. OS, power envelope, cooling, and software support all differ. CUDA cores, AMD compute units, and Apple GPU cores are counted differently, so core counts cannot be directly converted into performance ratios. Differences in memory bandwidth also do not directly translate to differences in inference speed.
That said, you can choose by use case. If existing x86 apps and peripherals are your priority, AMD-based Windows machines are a natural starting point for comparison. If your creative and AI work lives entirely in macOS, look at the M5 Max. For readers who want to use CUDA- and RTX-compatible creative software, development tools, and local AI on Windows, RTX Spark is a new option with a large shared memory pool. Qualcomm is also Windows on Arm, but you need to judge the NPU's INT8 TOPS, GPU processing, and software support separately.
Even With CUDA Running, Arm Compatibility Still Needs Checking
Native CUDA support matters a great deal to readers who rely on NVIDIA's existing software assets. RTX Spark has RT cores and Tensor cores and also supports DLSS 5, Reflex 2, Studio Driver, and Game Ready Driver. However, GPU-side support alone does not mean everything on the PC will run.
Windows 11 on Arm can translate and run x86 and x64 apps through Prism. But this applies only to user-mode apps; kernel-mode drivers and similar components must be compiled for Arm64. The more you depend on external devices, older business software, or specific development environments, the more you should check driver and native support before looking at the GPU spec sheet.
The same goes for games. On August 25, NVIDIA announced that EA, Embark, and Ubisoft are joining in supporting RTX Spark, citing titles such as Apex Legends and ARC Raiders as examples. For EA's Javelin Anticheat, it is working with EA on native support. Whether anti-cheat runs on Arm64 can determine whether a game launches at all, before graphics settings even come into play. Word that work is underway does not mean support is complete for every game.
For example, even if CUDA workloads run, any plugins or peripherals you use before and after that step that lack support can become obstacles when moving your work over. Check that your whole software setup can get the job done.
Real-World Performance to Check on the October Products
The up-to-1 PFLOPS figure NVIDIA cites is AI compute performance in the low-precision FP4 format. Both the conditions of the calculation and the circuitry that handles it differ from standard FP32 performance and from the 80 TOPS at INT8 shown for Qualcomm's NPU. You cannot rank AI speed on the magnitude of these numbers alone.
If you are buying for local AI, the deciding factors are generation speed and power consumption when running the same model with the same quantization settings and context length. A model fitting in memory is a separate condition from being able to work without waiting. For games, beyond whether the titles you usually play launch, it helps to compare with resolution and frame generation settings matched.
For people who want to combine CUDA-based creative work and large-capacity local AI in one machine, RTX Spark adds to the list of portable Windows candidates. Once pricing and real-app speeds become clear with the October products, and if your own software and peripherals are supported, the path to expanding into local AI while keeping your usual creative environment becomes much more concrete.