NVIDIA RTX Spark Driver Reveals Two Blackwell GPU Variants With Up to 6,144 CUDA Cores

NVIDIA appears to be preparing at least 2 RTX Spark configurations, with its first Windows 11 Arm64 driver revealing Blackwell GPUs containing 6,144 and 5,120 CUDA cores. The discovery provides the clearest evidence yet that RTX Spark will extend beyond the flagship configuration announced during Computex 2026.

NVIDIA published driver version 616.00 as an RTX Spark Developer Preview, allowing developers to begin porting applications, games and artificial intelligence tools to the new Windows on Arm platform before compatible hardware becomes widely available. The software is intended for development rather than consumer systems and currently includes several known performance and stability limitations.

After examining the driver package, Windows Latest discovered an installation file named nv_surface_woa.inf containing multiple RTX Spark N1X device identifiers. One entry describes a Blackwell RTX GPU with 6,144 CUDA cores, while 2 additional identifiers reference a second configuration with 5,120 CUDA cores. The file also includes entries for an NVIDIA NPU and an unidentified NVIDIA desktop device.

The 6,144 core configuration was already confirmed through NVIDIA and Microsoft’s original RTX Spark announcement. The newly discovered 5,120 core version is therefore the more significant detail, suggesting NVIDIA and its hardware partners will offer multiple performance tiers across laptops and compact desktop computers.

Assuming each Blackwell Streaming Multiprocessor contains 128 CUDA cores, the flagship configuration would use 48 Streaming Multiprocessors, while the reduced version would contain 40. However, the driver does not reveal GPU clock speeds, power limits, memory bandwidth or final system configurations, all of which will have a major influence on real performance.

The driver also does not confirm the CPU paired with either GPU. NVIDIA officially lists RTX Spark with up to 20 Arm CPU cores, but reports suggest the 5,120 core GPU could eventually be combined with an 18 core processor. That possible pairing remains speculation and should not be treated as a confirmed driver discovery.

RTX Spark combines its Blackwell GPU and Grace CPU through NVIDIA NVLink C2C, with platform configurations offering up to 128 GB of unified memory and 1 petaflop of artificial intelligence performance. Microsoft has also optimized Windows scheduling, memory management, power control and Prism emulation specifically for the architecture.

The platform is designed for artificial intelligence development, content creation and gaming. Microsoft has confirmed support for titles including Alan Wake 2, PRAGMATA, League of Legends, Valorant and PUBG Battlegrounds, while NVIDIA is bringing technologies such as CUDA, DLSS, Reflex, TensorRT and fifth generation Tensor Cores into slim laptops and compact desktop systems.

The first native RTX Spark Windows on Arm driver, also introduced CUDA 13.4 development support. The newly uncovered configuration confirms that NVIDIA is building a broader product family rather than relying on a single flagship design.

CUDA core counts alone cannot accurately predict gaming or creator performance. The 6,144 core total matches the desktop GeForce RTX 5070 and the older laptop GeForce RTX 3080, while 5,120 cores matches the laptop GeForce RTX 3070. These products use different architectures, frequencies, memory systems and power limits, making direct comparisons unreliable. NVIDIA itself warns that CUDA core counts should not be compared directly across different GPU generations.

The second configuration is important because it gives manufacturers more flexibility when designing RTX Spark systems. Premium creator laptops may use the complete 6,144 core GPU, while thinner or more affordable devices could adopt the 5,120 core version to reduce power consumption, cooling requirements and production costs.

The real question is how aggressively NVIDIA will limit the smaller configuration. Unified memory capacity, memory bandwidth and sustained power will matter more than the raw CUDA core difference. A well cooled 5,120 core system could outperform a heavily restricted 6,144 core laptop during extended gaming, rendering or artificial intelligence workloads.

Would you choose the 5,120 core RTX Spark model for better efficiency, or pay more for the complete 6,144 core Blackwell GPU?