Nvidia RTX Spark targets local AI laptops with 128GB unified memory

Nvidia’s RTX Spark platform is heading into laptops and compact desktops with a huge shared-memory pool aimed at running AI models locally. Nvidia rates the platform at up to 1 petaflop of FP4 AI performance, with 6,144 CUDA cores and up to 128GB of unified memory.¹

Why RTX Spark matters

The immediate appeal is model capacity. The source report says RTX Spark systems should be able to run language models of roughly 120 billion parameters locally, with a context window of up to one million tokens.² That does not automatically make every RTX Spark machine faster than a workstation with a discrete professional GPU, but it does give the platform more room for models that would otherwise spill into the cloud or hit graphics-memory limits.

Nvidia’s own product material positions RTX Spark as a platform for creators, developers and gamers, rather than a single laptop model.¹ Partner systems are expected from Asus, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and Gigabyte also named as future partners.²

The Dell comparison

The comparison that prompted the report is Dell’s Pro Max 18 Plus. Its top configuration reportedly pairs an RTX PRO 5000 with 128GB of system memory, but only 24GB of graphics memory is available directly to the GPU.² RTX Spark’s unified memory could therefore accommodate larger local models, although the source also notes that Dell’s graphics chip has higher memory bandwidth and more CUDA cores.

That distinction matters. Parameter capacity is not the same thing as inference speed, and RTX Spark will need comparable real-world token-generation results to prove that it can displace Dell’s workstation-class lead. Power efficiency could become another advantage: the source report says an RTX Spark chip may operate around 45W to 80W, while the Dell system is a much heavier desktop-replacement machine.²

Nvidia’s official announcement also frames the platform around on-device agents, CUDA, TensorRT, RTX graphics and Windows-native security features.¹ In practical terms, the pitch is a portable PC that can run demanding AI workloads without sending every prompt and file to a remote service.

What remains to be tested

The headline specifications are ambitious, but independent testing will decide how useful they are outside Nvidia’s preferred FP4 and sparsity scenarios. Laptop cooling, sustained power limits, software support and the final prices will all shape the experience. The first serious comparisons should measure the same models on the same prompts, rather than relying on peak TOPS or memory capacity alone.

Our opinion

RTX Spark looks more interesting as a local-computing platform than as another graphics badge. Giving laptops access to 128GB of unified memory could remove a genuine barrier for developers and creators who want to run large models without a workstation. But Nvidia still has to prove that its compact systems can turn that capacity into fast, quiet and sensibly priced machines. Until independent benchmarks arrive, RTX Spark is a promising challenge to Dell’s local-AI crown, not a confirmed winner.