Enthusiast reverse-engineers the NPU inside Nvidia's RTX Spark

Nvidia's RTX Spark (GB10) contains a 55 TOPS NPU for Copilot+ that the company barely documents. Firmware and driver analysis shows it is a standalone DLA-derived block with RISC-V cores, not a GPU virtualization, and ResNet-50 runs correctly on it in both FP16 and INT8.

• The NPU delivers 55 TOPS and is reserved for Copilot+ features
• It has its own registers, a 430–1562 MHz clock and separate power rail
• The driver identifies it as NVIDIA_DEV.DLA with NPU3 architecture
• ResNet-50 runs in INT8 but is no faster than in FP16