Nvidia RTX Spark NPU Runs Neural Nets at 55 TOPS Claim Despite Barely Any Windows Support
Updated
Updated · XDA Developers · Oct 24
Nvidia RTX Spark NPU Runs Neural Nets at 55 TOPS Claim Despite Barely Any Windows Support
2 articles · Updated · XDA Developers · Oct 24
Summary
Testing on HP's OmniBook Ultra 16 found RTX Spark's NPU is a physical accelerator—not a GPU front end—with its own registers, clocks, power rail and DLA-linked firmware, and it executed neural networks while the GPU stayed idle.
Beta Nvidia tools let the chip run FP16 and even unsupported INT8 workloads, including ResNet-50, but only through a fragile hybrid mode that submits DLA jobs through CUDA on the GPU.
Performance reached about 6 TFLOPS or TOPS at roughly 2.5-3 W on the NPU rail, making it faster than the CPU on tested models but still slower—and often less efficient overall once GPU overhead is counted—than Nvidia's own GPU.
INT8 worked only by bypassing TensorRT, delivered little to no speedup over FP16, and left Nvidia's advertised 50-55 TOPS unexplained; the company said an upcoming white paper will address the figure.
Windows support remains thin: DirectML and standard third-party interfaces cannot use the NPU, Copilot+ features are not yet available, and firmware evidence suggests the related GB10 platform may carry a disabled DLA block mainly added to satisfy Microsoft's Copilot+ requirement.