NVIDIA's RTX Spark Brings a Petaflop of AI Compute to the Windows PC

NVIDIA has launched RTX Spark, a chip that delivers 1 petaflop of AI performance to Windows laptops and small-form-factor desktops. The company announced the product at GTC Taipei, developed in collaboration with MediaTek and optimized with Microsoft, as The Silicon Review reported in June. First devices from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI will arrive in fall 2026.

The headline number is a petaflop of FP4 compute. But the architecture matters more than the peak throughput.

RTX Spark integrates a Blackwell RTX GPU with up to 6,144 CUDA cores and a 20-core Grace CPU built on Arm architecture, connected via NVLink-C2C at 600GB/s of bandwidth. The platform supports up to 128GB of unified LPDDR5X memory. NVIDIA says this configuration can run large language models with up to 120 billion parameters locally, with a context length of 1 million tokens.

That is a concrete capability claim worth testing. A 120B-parameter model at 4-bit quantization fits in roughly 60GB of memory, leaving headroom for context and application state. The 1-million-token context window suggests the memory bandwidth and unified memory architecture can sustain attention over very long sequences. For AI builders, this changes what “local inference” means. Today, running a 70B model on a consumer GPU requires careful quantization and often offloading layers to system RAM. RTX Spark claims to handle a model nearly twice that size entirely on-chip.

Microsoft has optimized Windows for the chip. The operating system now includes workload profile scheduling (WPS) and the Microsoft Power and Thermal Framework (MPTF) to balance performance, battery life, and heat. The Prism emulator for x86 apps has been tuned for the RTX Spark microarchitecture. These are not trivial software changes. They indicate that Microsoft sees RTX Spark as a platform shift, not a peripheral accessory.

NVIDIA and Microsoft are also building a native Windows platform for local AI agents. NVIDIA is bringing OpenShell to Windows using new security and containment primitives. Hermes Agent and OpenClaw will integrate OpenShell into their Windows applications, enabling secure on-device agent execution. This is the piece that most directly challenges Qualcomm and its Snapdragon X Elite lineup, which has been positioning itself as the Windows-on-Arm AI PC platform. Qualcomm’s NPU delivers roughly 45 TOPS. RTX Spark delivers 1,000 TOPS at FP4. The gap is two orders of magnitude.

Adobe is rebuilding Photoshop and Premiere from the ground up for RTX Spark, promising up to 2x AI and graphics performance. Native Arm apps also include Blender, DaVinci Resolve, Maxon Cinema4D, and CapCut. NVIDIA claims the chip can render 90GB 3D scenes, edit 12K video, generate 4K AI video, and play AAA games at 1440p exceeding 100 FPS with ray tracing and DLSS enabled.

The gaming claim is worth scrutiny. A chip with 6,144 CUDA cores and Blackwell architecture can certainly deliver competitive rasterization performance. But the unified memory architecture and the Arm CPU mean that game compatibility depends on Microsoft’s Prism emulator and native Arm builds. Early Snapdragon X Elite devices showed that x86 emulation works well for productivity apps but can introduce stutter in games. NVIDIA has the advantage of its own GPU driver stack and DLSS, which can compensate for emulation overhead. Still, the gaming experience will vary by title.

The bigger strategic question is what RTX Spark means for the PC processor market. NVIDIA has been a GPU supplier to PC OEMs for decades. RTX Spark is a full processor: GPU, CPU, memory, and interconnect on a single platform. It competes directly with Intel’s Core Ultra, AMD’s Ryzen AI, and Qualcomm’s Snapdragon X Elite. Each of those platforms includes an NPU for AI acceleration, but none delivers a petaflop of AI compute. Intel’s Lunar Lake NPU peaks at roughly 45 TOPS. AMD’s Ryzen AI 300 series NPU reaches 50 TOPS. Qualcomm’s Hexagon NPU in the Snapdragon X Elite hits 45 TOPS. RTX Spark is not incrementally better. It is a different class of device.

The collaboration with MediaTek is significant. MediaTek brings modem and connectivity expertise, plus established relationships with PC OEMs in the Asia-Pacific market. NVIDIA gets a partner that can help scale the platform beyond the high-end gaming and creator segment that its discrete GPUs already serve. MediaTek gets access to NVIDIA’s GPU and AI IP. The arrangement mirrors the Nintendo Switch partnership, where MediaTek also played a role in the custom Tegra chip.

First RTX Spark devices include the Microsoft Surface Laptop Ultra, ASUS ProArt, Dell XPS 16 Creator Edition, HP OmniBook, Lenovo Yoga Pro 9n, and MSI Prestige N16 Flip AI+. Acer and GIGABYTE will follow later. The OEM lineup is broad, covering consumer, creator, and enterprise segments. That breadth suggests NVIDIA is not targeting a niche. It is aiming for volume.

The open question is price. NVIDIA has not disclosed pricing for RTX Spark devices. The chip uses LPDDR5X memory, which is more expensive than standard DDR5. The Grace CPU and Blackwell GPU are both large dies. The NVLink-C2C interconnect adds complexity. If RTX Spark devices land at prices comparable to high-end Intel or AMD laptops, they will sell. If they carry a premium comparable to NVIDIA’s workstation GPUs, they will remain a creator and developer niche.

For AI builders, RTX Spark changes the local compute calculus. Today, running a 70B model on a laptop requires cloud connectivity or a desktop with multiple GPUs. RTX Spark claims to do it on a single chip in a laptop form factor. If the claim holds, it enables a new class of local AI applications: agents that run entirely on-device, with no latency from network calls and no data leaving the machine. Adobe’s full rebuild of Photoshop and Premiere suggests that software vendors are betting on this future.

The devices arrive in fall 2026. The silicon is real. The software stack is being built. The question is whether the PC market is ready for a chip that redefines what a laptop can do, or whether RTX Spark remains a tool for the builders who already know they need it.