In an Oct 16 technical blog post, NVIDIA published developer-workload benchmarks for RTX Spark laptops and compact desktops, which pair a Grace CPU with a Blackwell GPU in the GB10 system-on-a-chip. NVIDIA says the systems carry up to 128GB of unified memory and 1 petaflop of FP4 compute, enough to run coding models such as Qwen 3.8 Flash Next locally. Against an Intel Core Ultra X9 388H (Panther Lake 12Xe) normalized to 1.0, NVIDIA reports 4.6x on coding agents (AA-AgentPerf-local with Qwen3.8-27B), 2.7x on LoRA fine-tuning (Unsloth, Llama-3.1-8B) and 1.8x on Linux kernel compilation using Phoronix Test Suite. The post also describes a Windows Developer Config built with Microsoft, Windows Subsystem for Linux support, and Arm-native tooling including a PyTorch-CUDA build and more than 2,000 Conda packages.
Our takeThese are vendor-run numbers on three hand-picked workloads, normalized to a single Intel part, with no memory-bandwidth or thermal data published. The coding-agent result depends heavily on which model and quantization fit in 128GB, so it says more about capacity than about per-token speed. The claim worth checking independently is the 4.6x, once the machines ship in volume.