NVIDIA's RTX Spark, an ARM-based chip developed with MediaTek and previously known as N1X, has surfaced in new Geekbench 7 leaks showing improved performance compared to earlier disappointing results. Two RTX Spark laptops with 64GB of unified memory were tested in a "Balanced" power profile: one version with an 18-core CPU and one with a 20-core CPU. The 20-core model scored 2,570 in single-core and 23,126 in multi-core tests, while the 18-core version scored 2,541 and 21,776 respectively. By comparison, Apple's 14-core M4 Max posted a single-core score of 3,404 and a multi-core score of 25,957.

Based on these figures, the 20-core RTX Spark trails the M4 Max by about 11 percent in multi-core performance but by roughly 24.5 percent in single-core performance. Between the two RTX Spark variants, the 20-core CPU is about 1.14 percent faster in single-core tests and 6.2 percent faster in multi-core tests than the 18-core version. According to a developer driver for Windows 11 ARM64, the two RTX Spark configurations also differ on the GPU side: the slower CPU variant pairs with 5,120 CUDA cores, while the faster variant includes 6,144 CUDA cores on the Blackwell architecture. Notably, TechPowerUp reports that both the 5,120-core and 6,144-core CUDA configurations are paired with the same 20-core Arm CPU rather than separate 18-core and 20-core CPU variants, a detail that differs from the Geekbench leak's characterization of two distinct CPU core counts.

Separately, NVIDIA has expanded software support for the RTX Spark platform by bringing its RTX Neural Texture Compression (NTC) technology to Windows-on-Arm. NTC uses small neural networks to generate texture detail rather than relying on traditional block-compressed formats like BC5, BC6, or BC7, which are typically applied in 4x4 pixel blocks. NVIDIA has demonstrated that NTC can reduce VRAM usage substantially — by up to seven times in one case, and by as much as 85 percent in earlier demonstrations — including a specific scene where usage dropped from 6.5 GB to just 970 MB. The technique also enables low-bitrate compression that can unlock additional levels of detail, delivering up to 16 times more texels while using similar storage as standard compression methods, though viewers must move close to an object to notice degraded areas.

The NTC SDK, previously available only for Windows 11 on x86 platforms, is now available for ARM64 ahead of the expected release of RTX Spark hardware and related notebooks later this year. NVIDIA has also shipped an initial preview build of its CUDA toolkit as a native Windows-on-Arm software package, alongside the first Windows 11 ARM64 driver for the platform, suggesting that RTX Spark laptops could ship soon. Despite the SDK's availability, no shipping game currently supports NTC, and it remains limited to samples and technical demonstrations. One implementation of the compression method uses fused backpropagation to allow per-material optimization at resolutions up to 8,192 by 8,192, with a compressor capable of processing a 9-channel, 4K material texture set in one to fifteen minutes on an RTX 4090.

Because RTX Spark uses a unified memory architecture that can be configured with up to 128GB of LPDDR5X RAM, NTC's VRAM-saving benefits may be less critical on this platform than on dedicated GPUs, even though 32GB configurations are expected to be more common among buyers. Still, the framerate improvements NTC can provide would be welcome, as RTX Spark has already been shown running titles such as Alan Wake 2 and PRAGMATA. NVIDIA has also gained early software backing from Japan's SEGA, which has released AAA titles optimized for RTX Spark and is dedicating additional resources to support the platform's ecosystem. Some early performance figures have also emerged from Microsoft Surface Laptop Ultra prototypes, though these results are described as subject to change given the unfinished state of current drivers.