Nvidia RTX Spark: The Response to Apple Silicon and Snapdragon Chips

Nvidia RTX Spark: A New Era for Local AI on Windows

Nvidia recently made waves by unveiling RTX Spark at the GTC Taipei event. This innovative chip is designed for Windows laptops and stands out with its impressive specifications, including up to 128 GB of unified memory and a computing capacity reaching 1 petaflop in FP4. By directly targeting Apple Silicon and Qualcomm Snapdragon, Nvidia is making its entry into the Windows laptop market.

An Architecture Designed for AI

The RTX Spark chip is built to run AI agents locally, a breakthrough made possible by advanced security tools like OpenShell Runtime. This technology ensures the isolation of agents and protects user privacy. The first devices equipped with this chip, produced by giants such as ASUS, Dell, HP, Lenovo, and Microsoft Surface, are expected to arrive in the fall of 2026.

RTX Spark represents Nvidia's first step into the realm of Windows laptops, offering a Windows version of the already known DGX Spark chip. The hardware is based on the same GB10 Grace Blackwell Superchip that powers the DGX Spark, but it is aimed at a different audience. Rather than targeting Linux workstations for AI developers, RTX Spark is designed for consumers using Windows laptops and compact desktops.

Cutting-Edge Performance

The RTX Spark chip integrates a Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores. The Arm-based Grace CPU, developed with the help of MediaTek, features 20 cores and is connected to the GPU via NVLink-C2C. The shared memory reaches 128 GB, optimizing performance for AI tasks. The claimed peak of 1 petaflop refers to FP4 precision with sparsity, a theoretical scenario according to Nvidia's specifications. The GPU performance is said to be close to a GeForce RTX 5070 Laptop GPU depending on the workload, Nvidia claims.

Comparison with Apple and Qualcomm

Following Apple's strategy with its M-series chips, Nvidia offers an architecture that integrates CPU, GPU, and memory controller into a single package. Apple's M4 Max also provides up to 128 GB of unified memory with a bandwidth of 546 GB/s, but its Neural Engine peaks at 38 TOPS (INT8). In comparison, RTX Spark claims around 1,000 TOPS, although this is in FP4 with sparsity, so the conditions are very different. Nevertheless, the gap in raw AI computation is significant. Nvidia's real advantage lies in its CUDA stack, including TensorRT and RTX, which operates natively.

Qualcomm has also invested in Windows laptops on Arm with the Snapdragon X Elite in 2024, followed in September 2025 by the X2 Elite, boosting performance to 80 TOPS across 18 Oryon cores. These chips are built around Microsoft's Copilot+ features, not for local inference with multi-billion parameter models. Traditional x86 platforms from Intel and AMD continue to rely on separate CPU and GPU memory with much smaller NPUs.

Security and New Applications

Nvidia emphasizes the security of AI agents with enhanced Windows components for identity management and agent isolation. Nvidia OpenShell Runtime plays a key role in directing requests to local or cloud models while protecting personal data. Projects like Hermes Agent and OpenClaw have already integrated these features.

Meanwhile, Adobe plans to reinvent Photoshop and Premiere to leverage the capabilities of modern GPUs. On RTX Spark, these applications are expected to benefit from significant acceleration, with workflows potentially twice as fast.

A Powerful Workstation

Nvidia also introduced the DGX Station for Windows, equipped with the GB300 Grace Blackwell Ultra Desktop Superchip. With shared memory of up to 748 GB and a performance of 20 petaflops in FP4, this station is capable of handling AI models of unprecedented complexity. Its availability is expected in T4 2026.

Devices equipped with RTX Spark will be available starting in the fall of 2026, marking a significant milestone in the evolution of Windows laptops towards enhanced local AI capabilities.