A fresh look at an official Nvidia graphics driver suggests the company is preparing a second, scaled-down chip configuration for its newly announced ARM-based RTX Spark platform. The silicon, unveiled on May 31 alongside the first wave of compatible laptops, was initially detailed in just one Blackwell GPU variant featuring 6,144 CUDA cores. New driver evidence now points to a lower-tier model with 5,120 CUDA cores.
Driver package reveals a second configuration
The information surfaced inside driver version 616.00, which is intended for Microsoft’s RTX Spark Dev Box. Within the package, an ‘nv_surface_woa.inf’ file contains strings referencing two distinct graphics setups. One matches the previously disclosed 6,144 CUDA-core Blackwell GPU, while the other identifies a 5,120 CUDA-core variant. No further technical differentiators have been confirmed for the smaller configuration.
At this stage, it remains unclear how the core count reduction will affect other platform characteristics. Nvidia’s published ceiling for the RTX Spark chip includes up to a 20-core Grace CPU, up to 1 Petaflop of AI performance, and as much as 128 GB of unified memory. It is plausible that CUDA core count is the sole differentiator between the two tiers, a segmentation strategy reminiscent of Apple’s approach with its ‘M’ family of processors.
Hints of a desktop-class device and dedicated NPU
Beyond the GPU variants, the same driver package also makes reference to an “NVIDIA Desktop Device” and an “Nvidia NPU.” The desktop-related entry may be connected to the Surface RTX Spark Dev Box, though the driver itself offers few concrete clues about a standalone desktop product or a dedicated neural processing unit.
Additional pre-release benchmarks have recently placed the RTX Spark Arm chip in a Surface Laptop Ultra test system running a 20-core CPU. In those Cinebench results, the silicon proved competitive with AMD’s Ryzen AI Max+ 395 and Intel’s Core Ultra X9 388H, while falling behind Qualcomm’s Snapdragon X2 Elite and Apple’s M5 Max in both single- and multi-core workloads.