NVIDIA has unveiled the RTX Spark, a new Arm-based chip designed to run AI agents, complex workloads and high-end gaming on thin laptops and compact desktops.
The platform is the result of a multi-year collaboration between NVIDIA and Microsoft to bring NVIDIA’s Grace and Blackwell technologies into a new generation of Windows PCs. The two companies are also jointly developing the software foundation for on-device AI, combining NVIDIA’s AI stack with new Windows capabilities created to support secure, locally executed AI agents.
Running AI agents locally instead of using the cloud requires an advanced chip. The RTX Spark combines an NVIDIA Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores, linked by NVLink-C2C to a 20-core Grace CPU developed with MediaTek.
NVIDIA says RTX Spark can deliver up to 1 petaflop of AI performance and support as much as 128GB of unified memory, a robust configuration aimed at workloads that have typically required workstations or cloud systems.
“NVIDIA recognizes that AI will need to move to the edge as data center costs rise. Yet this announcement risks overpromising the capabilities of local agents given the memory intensity and CPU requirements of agentic experiences,” Brendan Burke, Research Director at The Futurum Group, told Techstrong.it.
“MediaTek has conventionally made lower-end mobile chips so this collaboration presents a step up in compute intensity that may not succeed in its first generation. Potential bottlenecks in GPU orchestration will need to be independently tested.”
However, Burke noted, “with NVIDIA’s spectacular track record in developer ecosystems, the product may soon overcome its technical hurdles to become a catalyst for mass adoption of on-device AI.”
Controlling AI Agents
Microsoft is adding Windows security and containment capabilities, while NVIDIA is providing OpenShell, a runtime intended to control what actions agents can take and when queries should remain local rather than being sent to cloud models.
For enterprises and developers, part of the significance is privacy and control. NVIDIA says RTX Spark systems can run 120-billion-parameter LLMs locally with context windows of up to 1 million tokens. The company laid out some impressive potential use cases: users will be able to render 90GB-plus 3D scenes, generate 4K AI video and play AAA games at 1440p above 100 frames per second.
Adobe is reworking Photoshop and Premiere for RTX Spark, with NVIDIA claiming up to 2x faster AI and effects performance. Premiere is expected to use the chip’s unified memory, Blackwell GPU, and TensorRT software for real-time editing and color correction, while Photoshop will gain a GPU-accelerated compositing engine for advanced HDR graphics editing.
Systems Due This Fall
The first RTX Spark systems are expected this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and GIGABYTE to follow. NVIDIA has said more than 30 laptops and 10 desktops are planned. Systems will include Windows laptops as thin as 14 millimeters and as light as about three pounds, with 14- to 16-inch displays and premium chassis designs.
For example, Microsoft previewed the Surface Laptop Ultra, an RTX Spark-based system expected to offer up to 128GB of unified memory, a 15-inch mini-LED PixelSense Ultra display with 2,000 nits peak HDR brightness, 2880 x 1920 resolution and 262 pixels per inch.
For the PC market, NVIDIA’s new chip introduces new competition. RTX Spark puts NVIDIA into a PC processor market long dominated by Intel, AMD, Qualcomm and Apple. It also expands Windows on Arm into higher-performance systems, especially where developers need CUDA support and GPU-heavy local AI.
“This will put pressure on the incumbents, like Intel and AMD, but they still command the vast majority of the AI PC market, and NVIDIA will likely see a slow growth in this space relative to the significant AI PC market expansion,” Jack Gold, president of J. Gold Associates, told Techstrong.it.
Workstations Run 1 Trillion Parameters Locally
NVIDIA is extending the same Windows strategy to enterprise workstations. The company also announced DGX Station for Windows, a deskside AI system powered by the GB300 Grace Blackwell Ultra Desktop Superchip.
That system pairs a 72-core Grace CPU with a Blackwell Ultra GPU, supports up to 748GB of coherent memory and delivers up to 20 petaflops of FP4 AI performance. NVIDIA says it can run models with up to 1 trillion parameters locally and will arrive in Q4.
Originally published by Techstrong.IT. Republished with attribution.
