NVIDIA RTX Spark
NVIDIA is bringing more of its AI and graphics technology into Windows laptops with RTX Spark, a platform that combines an Arm-based Grace CPU, Blackwell graphics and up to 128 GB of unified memory. Laptops start at $2,599 and become available on 16 October, while compact desktops are scheduled for November. We don’t have Indian prices yet but a close approximation would mean that these laptops start at Rs 3,00,000 and we should get more details in the coming week.
With the RTX Spark, the proposition stretches across gaming, content creation and software development, but local AI provides the most significant shift. A large pool of memory accessible to both the CPU and GPU gives these systems room for models and creative projects that would otherwise exceed a laptop GPU’s dedicated video memory. NVIDIA is also building out the surrounding software, from coding agents and application controls to video tools and AI-powered game characters.
That makes RTX Spark a broader proposition than a just a laptop with an AI assistant attached. NVIDIA wants the same machine to run demanding applications, generate content and host agents that can work across files and software, with cloud services available when needed. If this isn’t the truest sense of the phrase “AI PC”, then we don’t know what is.
At the top of the RTX Spark N1X range is a 20-core Grace CPU, developed with MediaTek, alongside a Blackwell GPU with 6,144 CUDA cores. The CPU combines 10 performance cores with 10 efficiency cores. NVIDIA also lists an 18-core CPU configuration paired with 5,120 GPU cores.
| Platform feature | Specification |
| CPU | Up to 20 Grace cores, comprising 10 performance and 10 efficiency cores |
| GPU | Blackwell graphics with up to 6,144 CUDA cores |
| AI compute | Up to 1 petaflop at NVFP4 precision |
| Unified memory | Up to 128 GB LPDDR5X |
| Memory bandwidth | 300 GB/s |
| CPU-to-GPU interconnect | NVLink C2C, rated at 600 GB/s |
| Platform connectivity | PCIe Gen 5, up to four 40 Gbps USB4 ports and MediaTek Wi-Fi 7 |
| Display interfaces | HDMI 2.1a and DisplayPort 2.0 |
The memory bandwidth and interconnect figures describe different parts of the architecture. LPDDR5X supplies the shared memory pool at 300 GB/s, while NVLink C2C connects the CPU and GPU tiles with a coherent link rated at 600 GB/s. Unlike the DGX Spark boxes, the Spark laptops don’t feature ConnectX-7 ports, at least none of the units we saw do. So daisy chaining multiple RTX Spark machines isn’t going to be as easy as the DGX machines.
For local AI, the practical benefit is capacity. A model must fit into the memory available to the inference software, alongside its working data. More memory can accommodate larger models, longer contexts or additional workloads. Creative applications can use that capacity for complex scenes and demanding projects too. The full 128GB is not a dedicated allocation for the GPU. Windows and other applications also need memory, so usable capacity varies with the configuration and workload. You can always tweak the allocation within the NVIDIA App, so you get a bit more control than typical laptops with high memory capacity.
Blackwell brings fifth-generation Tensor Cores, fourth-generation RT Cores and NVIDIA’s graphics and compute software stack. The headline 1-petaflop figure applies to low-precision NVFP4 AI computation; it is not a measure of general CPU performance or gaming speed. The video engine also supports 4:2:2 encoding and decoding, AV1 UHQ and higher H.264 decoding throughput.
Microsoft and NVIDIA have worked on scheduling, power management, graphics and memory handling for RTX Spark. Workload profile scheduling helps Windows distribute tasks across the CPU cores, while the Microsoft Power and Thermal Framework coordinates performance and thermal behaviour as workloads change. Windows also gains a higher, more adaptive limit on GPU-accessible system memory, alongside improvements to memory page handling. These changes help applications use the larger shared memory pool for AI inference and creative work.
Application compatibility remains central to any Arm-based Windows platform. Native Arm applications highlighted for RTX Spark include Blender, DaVinci Resolve, Affinity and Adobe Photoshop. Existing x86 applications can use Microsoft’s Prism emulation where supported. Native execution and GPU acceleration are separate considerations, however. An application’s supported version and features determine whether a particular task can use Blackwell, CUDA or another accelerated path. The presence of an RTX GPU does not automatically accelerate every application running through emulation.
NVIDIA’s agent strategy focuses on software that can take action through compatible tools. That could mean working with documents, managing a coding project, interacting with creative applications or coordinating household devices. RTX Spark is being positioned to run models such as the 125-billion-parameter Qwen3.8-Flash-Next mixture-of-experts model locally. Memory capacity works for fitting a model but response speed depends on its format, context and inference runtime.
Local execution can keep model processing on the device and reduce reliance on metered cloud inference. An agent that accesses an online service or cloud connector can still transmit information through that service, so the privacy implications depend on the whole workflow.
Cloud models remain part of the plan. Since frontier models are mostly going to be placed behind paywalls, there’s always the need to refer to them when the local models don’t match up. For such situations, NVIDIA’s NeMo Switchyard routes work between local and cloud models, using a verifier to determine whether the local response has completed the task. More demanding requests can then move to a cloud model. A beta is expected in Hermes Agent towards the end of October.
We got a taste of the Hermes harness and boy does it impress. We used it to run through more than 20 videos, transcribe them all, sift through them to identify key phrases, and then weave a logically sound paragraph using quotes from each video. Prima facie, you wouldn’t even be able to tell that the paragraph was put together like Frankenstein’s monster. And it did all of that in under 30 minutes. Of course, a few additional tools had to be downloaded along the way, and the Hermes agent was configured to seek permission for anything extending beyond its purview. That being said, this was the closest to getting a Jarvis-like (Iron Man reference) experience.
NVIDIA is also extending the ecosystem of application skills and Model Context Protocol (MCP) integrations. These give compatible agents routes into productivity software, development tools and creative applications, including NVIDIA’s own utilities.
Giving software access to files and applications requires controls over what it can do. NVIDIA is bringing OpenShell to Windows with Microsoft Execution Containers integration, combining agent policies with operating-system containment.
The proposed controls cover file access, applications, websites and credentials. Windows also supports agent identities that distinguish agent activity from actions performed directly by a user. Actual support depends on the Windows build and the agent’s integration. Hermes support for OpenShell, Microsoft Execution Containers and hybrid prompt routing is expected by the end of October.
For homes or teams with multiple computers, NVIDIA PAIR distributes inference across machines on a local network. The tool discovers and pairs devices, then routes work to available systems. It supports Ollama, LM Studio and, newly, llama.cpp.
PAIR reached general availability on 13 October and is open source under the Apache 2.0 licence. A compact RTX Spark desktop could therefore serve models to other devices around a home, rather than requiring each laptop to host the same model independently.
For developers, RTX Spark brings the Windows desktop, Linux tooling through Windows Subsystem for Linux and CUDA acceleration into one system. The aim is to let a developer write code, run a local coding agent, test a project and fine-tune supported models without repeatedly moving between a laptop and a cloud machine.
The software ecosystem includes PyTorch, llama.cpp, Ollama, LM Studio, vLLM, SGLang, Hugging Face and Unsloth. NVIDIA says more than 2,000 Conda packages are available and 97% of the top 15,000 Python packages are supported. NVIDIA is also making its libraries available through agent skills, helping compatible coding agents find and use CUDA-X capabilities. GitHub Copilot integration includes local AI through Windows ML, local-to-cloud prompt routing and Microsoft Execution Containers. Support through llama.cpp is described as coming later.
The wider developer roadmap includes DGX Station for Windows, scheduled for December with 20 petaflops of AI compute and 748GB of coherent memory. That system targets teams serving larger models and accelerated applications, rather than the portable role of an RTX Spark laptop.
NVIDIA lists more than 130 RTX-accelerated creative applications, spanning editing, rendering and generative AI. RTX Spark’s shared memory is intended to support large 3D scenes and local image and video models, while its video engine handles formats such as native 4:2:2 footage.
Adobe’s Photoshop, Premiere and Substance 3D are part of the platform push. NVIDIA claims up to twice the speed across selected AI, editing, colour and effects workflows, although gains will vary by application and task. For generative AI, NVIDIA says supported NVFP4 models can deliver 1.5 times the performance while being 40% smaller. Lower-precision model formats can reduce memory requirements, but output quality still depends on the individual model and its quantisation.
RTX Video Frame Generation and HDR are also coming to ComfyUI, extending the tools available for generated video. Separately, DLSS 4.5 Ray Reconstruction is scheduled to arrive in Blender 5.3 in November for RTX Spark and GeForce RTX hardware.
The updated NVIDIA Broadcast, available from 13 October for RTX Spark and GeForce RTX, adds RTX Video Super Resolution, RTX Video Frame Generation, Studio Voice profiles and an offline File Processor. MCP support also lets compatible agents control Broadcast, opening a route to automated media-processing workflows.
NVIDIA says RTX Spark supports hundreds of leading games at launch. Its gaming stack includes ray tracing, DLSS Super Resolution, Frame Generation, Multi Frame Generation, Ray Reconstruction, Reflex and G-SYNC. Compatibility also depends on the game and its supporting software. NVIDIA spoke about work involving KRAFTON, EA Javelin Anticheat, Activision and Riot, an important area for multiplayer titles on Windows on Arm. Support for individual games remains the relevant check.
The DLSS October update has two parts. A Performance model for Frame Generation entered beta in the NVIDIA app on 13 October for RTX Spark, RTX 40 Series and RTX 50 Series GPUs. A faster DLSS 5 model follows on 16 October for RTX Spark and RTX 50 Series GPUs, with NVIDIA claiming twice the model performance.
That claim concerns the neural model’s execution speed. It does not mean every supported game will deliver twice the overall frame rate. DLSS 5 is also scheduled for Hogwarts Legacy on 5 November and Where Winds Meet in late November, covering RTX Spark, GeForce RTX 50 Series and GeForce NOW.
NVIDIA’s gaming announcements extend to AI characters through ACE. Nemotron 3.5 Bolt E4B, scheduled for 23 October, supports configurable conversational and reasoning modes, with open training sets, weights and fine-tuning recipes. NVIDIA claims 1.5 times higher accuracy than the previous model on an internal composite of conversational-agent tasks. The intended uses include characters that respond to dialogue and game context, alongside community mods built around local AI.
RTX Remix is expanding too. A Minecraft Java mod kit will add path tracing and DLSS while supporting existing Java maps and upgraded particle and volumetric effects. A beta kit, sample map and high-resolution texture pack are scheduled for 23 October. DLSS 5 support for RTX Remix arrives on the same date.
RTX Spark laptops start at $2,599 and become available on 16 October. Compact desktops follow in November. Indian pricing and model-specific availability have not been made available since OEMs tend to play it safe and only reveal details after all models have completed BIS certification. However, we can approximate that the RTX Spark models will be priced starting Rs 3,00,000. For something that packs the same performance as the DGX Spark which costs Rs 7,00,000 currently, this is a good deal. It’s just the memory part that you’ll have to compromise on. Or not, since RTX Spark laptops will arrive with 128 GB RAM like the Surface Ultra. But that thing costs $5,899 and that’s going to be just as expensive as the DGX Spark box.
The platform brings together capabilities that have often required separate compromises: substantial GPU-accessible memory, RTX graphics, local AI software and a portable Windows form factor. Its usefulness will depend on the applications, games and models each system can run well, but the direction is clear. NVIDIA is making local AI a central part of the PC’s workload, alongside creation and gaming, and building the software around it to do more than hold a conversation.