NVIDIA RTX Spark: Windows AI PC Launch, Specs and Architecture
NVIDIA is preparing to enter the Windows PC processor market with RTX Spark, a unified computing platform that combines an Arm-based CPU, Blackwell GPU, high-bandwidth unified memory, and dedicated AI acceleration in a single SoC-oriented package.
At SIGGRAPH 2026, NVIDIA confirmed that the first RTX Spark systems are scheduled to launch in autumn 2026. The platform represents a significant departure from conventional Windows PC architectures, where the CPU and discrete GPU are separate devices connected through the system motherboard.
Instead, RTX Spark integrates the major compute components into a tightly coupled architecture designed for local AI, graphics, gaming, and general-purpose workloads.
🚀 RTX Spark Launch Timeline and OEM Support #
NVIDIA’s first-generation Windows PC platform is expected to debut across multiple form factors, with several major PC manufacturers preparing systems around the new architecture.
Initial hardware partners #
ASUS and MSI are expected to lead the first wave of RTX Spark products, followed by Acer and GIGABYTE.
Major Tier-1 OEMs, including Dell, HP, Lenovo, and Microsoft Surface, are also preparing systems based on the platform.
The expected product range extends from thin-and-light notebooks to compact desktop systems, suggesting that NVIDIA intends RTX Spark to compete across multiple segments rather than positioning it solely as a high-end workstation platform.
🧩 Two RTX Spark Configurations Emerge #
NVIDIA’s developer preview of the GeForce 616.00 graphics driver for Windows 11 on Arm reportedly provides additional information about the platform.
Analysis of the nv_surface_woa.inf package identifies two distinct hardware configurations.
| Configuration | CPU | GPU | CUDA Cores | Unified Memory |
|---|---|---|---|---|
| High-End | 20-core Arm CPU | Blackwell | 6,144 | Up to 128GB LPDDR5X |
| Standard | 18-core Arm CPU | Blackwell | 5,120 | Up to 128GB LPDDR5X |
The high-end configuration pairs a 20-core CPU with a Blackwell GPU containing 6,144 CUDA cores, broadly placing its GPU compute configuration in the vicinity of a desktop RTX 5070-class design.
The standard configuration reduces both CPU and GPU resources while retaining support for up to 128GB of LPDDR5X unified memory.
Unified memory as an AI advantage #
The unified-memory architecture is particularly relevant to local AI workloads.
Instead of maintaining separate CPU and GPU memory pools, RTX Spark allows both compute domains to access a shared high-capacity memory subsystem. This can reduce the need to explicitly move large model weights and tensors between system RAM and discrete GPU VRAM.
For local inference, that architecture can be useful when running models whose parameter counts exceed the practical VRAM capacity of conventional consumer GPUs.
🏗️ RTX Spark Architecture: A Unified Computing Design #
NVIDIA originally presented RTX Spark as a fundamental redesign of the PC architecture.
The platform combines CPU and GPU compute resources within a tightly integrated package and uses NVLink-C2C to provide a high-bandwidth connection between the processor complexes.
Silicon and package architecture #
The reported flagship configuration includes:
- Process: TSMC 3nm
- Transistor count: Approximately 70 billion
- CPU: 20-core Armv9-A implementation
- GPU: Blackwell architecture
- Interconnect: NVLink-C2C
- CPU-GPU bandwidth: Up to 600 GB/s bidirectional
- Unified memory: Up to 128GB LPDDR5X
The architectural objective is to make CPU and GPU resources behave more like components of a single compute fabric rather than independent processors connected through conventional PC interfaces.
CPU complex #
The CPU is reportedly co-developed with MediaTek and based on the Armv9-A architecture.
The 20-core configuration consists of:
- 10 Cortex-X925 cores operating at up to 4.1 GHz
- 10 Cortex-A725 cores operating at up to 2.86 GHz
- 1MB of L2 cache per core
- 32MB shared L3 cache
- 20 physical cores and 20 threads
This hybrid core arrangement provides a combination of high-performance and efficiency-oriented CPU resources while maintaining a relatively compact SoC footprint.
Blackwell GPU complex #
The flagship GPU configuration reportedly contains:
- 6,144 CUDA cores
- 48 fourth-generation RT Cores
- 192 Tensor Cores
- Up to 2,450 MHz GPU frequency
- DLSS 4.5
- Frame Generation
- NVIDIA Reflex
- G-SYNC
- Hardware AV1 encode/decode
- Up to 8K media processing
The combination of CUDA, RT, and Tensor acceleration gives RTX Spark a broader workload profile than a conventional Arm PC processor.
It can simultaneously target general-purpose applications, GPU compute, real-time ray tracing, AI inference, and gaming workloads using NVIDIA’s existing software ecosystem.
⚡ Memory, Storage and I/O #
Memory bandwidth is another important component of RTX Spark’s design.
The platform reportedly uses a 256-bit LPDDR5X interface operating at up to 8,533 MT/s, delivering approximately 273 GB/s of aggregate memory bandwidth.
Capacity can reach 128GB, providing significantly more memory than is typically available as dedicated VRAM in many consumer laptop GPUs.
Platform I/O #
The reported platform specification includes:
- Dual M.2 PCIe 4.0 storage interfaces
- Wi-Fi 7
- 10Gb Ethernet
- DisplayPort 2.1
- HDMI 2.1
- LPDDR5X unified memory
This combination gives RTX Spark enough I/O capability to function as both a high-performance notebook platform and a compact desktop compute system.
🤖 RTX Spark Targets Local AI #
Local AI is one of the most important workloads behind the RTX Spark architecture.
NVIDIA reportedly rates the platform at up to 1 PFLOPS of FP4 inference performance, while its large unified-memory configuration is designed to support local execution of models containing up to approximately 120 billion parameters.
The platform is also intended to support context windows of up to 1 million tokens, depending on the model and implementation.
CUDA and AI software compatibility #
RTX Spark is designed around NVIDIA’s established AI software ecosystem, including:
- CUDA
- TensorRT
- PyTorch
- Tensor Cores
- FP4 inference
- NVIDIA GPU acceleration libraries
This software compatibility could be one of the platform’s strongest advantages.
Developers already using CUDA-based workloads can potentially migrate applications to an Arm-based Windows system without abandoning NVIDIA’s broader GPU compute ecosystem.
The more challenging component will be native Windows on Arm application support, particularly for software that depends on x86-specific binaries, drivers, plugins, or low-level system integrations.
📊 Early Performance Positioning #
Preliminary leaked CineBench 2026 results reportedly place RTX Spark at approximately:
- 540 points single-core
- 5,771 points multi-core
These figures suggest that NVIDIA is not necessarily targeting absolute CPU performance leadership.
RTX Spark versus AMD #
The reported Ryzen AI Max+ 395 scores approximately 620 points in single-core and 6,700 points in multi-core testing, giving AMD an advantage in raw CPU throughput.
RTX Spark’s potential advantage instead lies in combining competitive CPU performance with a substantially more capable integrated Blackwell GPU and a large unified-memory pool.
RTX Spark versus Intel #
Against the Core Ultra X9 388H, reported RTX Spark scores are higher in both single-core and multi-core testing.
The comparison is particularly relevant because both platforms target premium Windows systems, although their architectural approaches differ considerably.
RTX Spark versus Apple Silicon #
The leaked results place RTX Spark in a broadly comparable overall performance tier to Apple’s 14-core M3 Max in these CPU benchmarks.
However, benchmark comparisons between these platforms should be treated cautiously because CPU performance represents only one portion of RTX Spark’s intended workload profile.
Its more significant differentiator is the combination of CPU, Blackwell GPU, Tensor acceleration, unified memory, and CUDA software compatibility.
💻 Compact Hardware and Thermal Design #
The integration of CPU and GPU resources into a unified package substantially changes motherboard design.
Traditional gaming laptops often require separate CPU and discrete GPU packages, dedicated VRAM, high-speed interconnects, and large power-delivery systems.
RTX Spark can eliminate much of this duplication.
Motherboard integration #
Reference motherboard designs reportedly show the RTX Spark SoC positioned at the center of the board, surrounded by unified LPDDR5X memory.
+-----------------------------------------------------------------------+
| ASUS ProArt RTX Spark Motherboard |
| |
| +---------------------------------------------------------------+ |
| | RTX Spark SoC | |
| | 20-Core CPU + Blackwell GPU + NVLink-C2C | |
| +---------------------------------------------------------------+ |
| |
| [M1] [M2] [M3] [M4] [M5] [M6] [M7] [M8] |
| 8x LPDDR5X Unified Memory Modules |
| |
| [ 12-Phase VRM ] [ M.2 PCIe NVMe #1 ] |
| [ M.2 PCIe NVMe #2 ] |
+-----------------------------------------------------------------------+
The elimination of conventional CPU and discrete-GPU sockets allows motherboard designers to reduce board area and simplify high-speed signal routing.
Thin laptops and mini PCs #
Reported RTX Spark systems can reach laptop chassis thicknesses of approximately 14 mm and weights around 1.36 kg.
Thermal solutions are expected to use dual-fan designs combined with multiple heat pipes.
The initial notebook range is expected to cover 14-inch and 16-inch systems, including configurations with tandem OLED displays and G-SYNC support.
Compact desktop systems are another natural fit because the integrated architecture reduces the physical volume normally required for a discrete GPU and its associated power and cooling hardware.
🔐 Windows on Arm and Local AI Security #
NVIDIA is also developing the software infrastructure required to make local AI agents practical on Windows on Arm.
The platform reportedly combines Microsoft security mechanisms with an NVIDIA runtime layer called OpenShell.
Dual-layer security architecture #
The proposed security model consists of two primary layers:
- Windows Security Primitives handle identity, process isolation, permissions, and operating-system-level policies.
- NVIDIA OpenShell Runtime provides explicit boundaries for local AI agents and controls how agents interact with local resources and external services.
This architecture addresses an important problem with autonomous local agents: an AI application may require access to files, applications, credentials, or network services to perform useful tasks.
Restricting those capabilities through explicit runtime permissions can reduce the risk of an agent accessing information beyond its intended scope.
Local versus cloud inference #
The runtime is also intended to manage the boundary between local and cloud AI execution.
When an operation can be handled locally, the system can use the RTX Spark compute stack. When cloud processing is required, sensitive information can potentially be sanitized before leaving the device.
This hybrid model could become increasingly important as AI assistants transition from passive chat interfaces toward autonomous agents capable of interacting with local applications and data.
🎮 Gaming and Software Ecosystem #
RTX Spark is not positioned solely as an AI workstation platform.
NVIDIA is bringing its established graphics ecosystem to Windows on Arm, including CUDA, ray tracing, DLSS, Reflex, and G-SYNC.
The company is targeting 100+ FPS at 1440p in modern AAA games under appropriate configurations.
Developer adoption #
More than 100 Windows software vendors have reportedly committed to native optimization for RTX Spark.
Native Arm support will be an important factor in determining whether the platform can compete effectively against x86 Windows systems.
While emulation can provide compatibility for legacy applications, native binaries are generally preferable for demanding workloads because they can avoid translation overhead and provide better access to platform-specific capabilities.
NVIDIA and SEGA partnership #
NVIDIA and SEGA have also expanded their partnership around RTX Spark.
The upcoming VIRTUA FIGHTER CROSSROADS, scheduled for 2027, is expected to receive Day-1 native optimization for the platform.
A growing library of native games and applications could help NVIDIA address one of the central challenges facing Windows on Arm: establishing a sufficiently broad software ecosystem to compete with mature x86 platforms.
🔭 RTX Spark’s Strategic Significance #
RTX Spark represents a significant expansion of NVIDIA’s role in the PC market.
Rather than supplying only the GPU, NVIDIA is moving toward controlling a larger portion of the compute architecture, combining an Arm CPU, Blackwell graphics, Tensor acceleration, unified memory, high-speed CPU-GPU interconnects, and an integrated software stack.
The most important differentiator may not be raw CPU benchmark performance. Instead, RTX Spark’s value proposition lies in combining large unified memory, high-performance GPU compute, local AI acceleration, CUDA compatibility, and compact system design within a single platform.
If NVIDIA can establish strong Windows on Arm application compatibility and deliver competitive pricing, RTX Spark could become an important new category of AI PC architecture.
The autumn 2026 launch will therefore test more than NVIDIA’s processor design. It will determine whether a GPU-centric company can successfully redefine the Windows PC around local AI, unified compute, and tightly integrated heterogeneous processing.