NVIDIA’s RTX Spark laptop platform is reaching customers through computer makers worldwide starting Oct. 16, with preorders already open, according to the company’s launch announcement. NVIDIA is pitching the systems for gaming and creative work as well as running AI workloads locally on a Windows PC. [1]
RTX Spark is built around a combined CPU-and-GPU superchip rather than a standalone graphics card. NVIDIA’s launch page describes a Grace CPU with up to 20 cores, a Blackwell RTX GPU with up to 6,144 CUDA cores, fifth-generation Tensor Cores, and an NVLink-C2C connection between the processor components. Microsoft says the platform supports up to 128 GB of unified memory and up to one petaflop of AI performance. Those figures are manufacturer specifications, not a single standardized performance result across every laptop configuration. [1] [2]
The large shared memory pool is central to NVIDIA and Microsoft’s local-AI pitch: it can let compatible applications allocate more memory to models and other GPU-heavy work without sending every task to a cloud service. The companies are also promoting Windows support for local agents and creator software. RTX graphics features and a selection of major games are scheduled to work on the platform at launch, although compatibility remains specific to each title and application. [1] [2]
Early tests show a trade-off
Windows Central tested a high-end, 20-core RTX Spark configuration before launch. Its reviewer found the GPU to be the system’s strongest component, while noting that single-core CPU results trailed some leading Apple and Qualcomm processors. The review also found that battery use rises under sustained, GPU-heavy workloads, even though lighter day-to-day use is less demanding. [3]
That makes the first systems most relevant to people who need substantial local GPU or AI capability in a portable Windows machine. The launch is a new hardware option, not evidence that every RTX Spark laptop will match the same benchmark, memory capacity, or battery runtime; those details depend on the specific OEM configuration and workload.