Nvidia RTX Spark laptops arrive with fast local AI and premium pricing

Nvidia RTX Spark laptops became available October 16 from ASUS, Dell, HP, Lenovo, Microsoft and MSI, according to Nvidia. The Windows systems target local AI work with configurations reaching 128GB of unified memory, but an independent review found the appeal comes with prices from about $2,600 to more than $6,000.

Nvidia says RTX Spark combines a Blackwell RTX GPU and Grace CPU, with top configurations offering 6,144 CUDA cores, 20 CPU cores and up to 1 petaflop of FP4 AI performance. Nvidia also says CUDA runs natively on the platform. Those specifications give the company a Windows notebook offering aimed at developers and creators who want to run models and AI-agent workloads on-device rather than send every request to a remote service.

The company’s claims about supported applications, gaming results, battery life and the largest models or creative projects the systems can handle have not been independently established by the material reviewed here. Nvidia did not disclose a single platform-wide price in its launch announcement.

Are Nvidia RTX Spark laptops worth the price for local AI?

For most laptop buyers, probably not. The Register’s hands-on assessment of HP’s OmniBook Ultra 16, an RTX Spark system, called the machines poor value at their reported price range despite finding the HP system difficult to beat among Windows notebooks for local AI. Its buyer case is narrower: AI enthusiasts, researchers and creative professionals who need substantial local model memory and prefer Nvidia’s CUDA software environment.

The publication tested the OmniBook on Windows with Llama.cpp v0.6.0, using five models from Meta, Google and Alibaba ranging from roughly 26 billion to 122 billion parameters. Its first chatbot-style workload processed a 256-token input and generated a 256-token response, measuring time to first token and tokens per second. The available report does not provide the underlying numerical benchmark tables, so it does not establish a specific performance ranking.

The hardware matters most where model weights and context consume significant memory. More physical memory does not mean all of it is immediately available for a model, however. The reviewed HP system initially divided its 128GB in half, assigning 64GB to the CPU and 64GB to the GPU, according to The Register. The allocation can be changed, but that default split is a practical constraint for buyers assessing model capacity on Windows.

Llama.cpp can extend into CPU-assigned memory when needed. Because the CPU and GPU use the same memory pool, The Register said the performance cost is less pronounced than it would be with a dedicated GPU. It also noted that model weights are only part of the requirement: large context windows consume memory too.

The top N1X configuration reviewed by The Register has 20 CPU cores and 6,144 CUDA cores, with an equal mix of Arm X925 and Cortex-A725 CPU cores. The publication put its memory bandwidth at about 300GB/s. Its tested HP model also has a 2,880-by-1,800 OLED touchscreen and weighs 4.89 pounds, though those design details are secondary to the platform’s memory and CUDA proposition.

RTX Spark is therefore a specialized premium notebook category, rather than a general-purpose value play. It should also not be confused with Nvidia’s Linux-based DGX Spark desktop platform, which Nvidia markets for always-on desktop agent workloads.

This story draws on original reporting from The Register.