Nvidia RTX Spark: 1-Petaflop Superchip Brings Local AI Agents to Windows PCs
Nvidia today announced RTX Spark, a new superchip that brings 1 petaflop of AI compute to Windows PCs, purpose-built for running personal AI agents locally. The chip combines a Blackwell RTX GPU (6,144 CUDA cores, fifth-gen Tensor Cores with FP4), a 20-core Grace CPU, and up to 128GB of unified memory, connected via NVLink-C2C. MediaTek contributed to the custom Arm-based CPU design for power efficiency.
Key Specs and Capabilities
- AI performance: 1 petaflop (FP4)
- GPU: Blackwell RTX with 6,144 CUDA cores
- CPU: 20-core NVIDIA Grace (Arm), co-designed with MediaTek
- Memory: up to 128GB unified memory
- Software stack: CUDA, RTX, DLSS, FP4, TensorRT, OptiX, Reflex, G-SYNC
RTX Spark can run 120B-parameter LLMs with up to 1 million tokens context locally, render 90GB+ 3D scenes, edit 12K 4:2:2 video, generate 4K AI video, and play AAA games at 1440p 100+ fps.
Windows-Native Agent Security
Nvidia and Microsoft are collaborating on new Windows security primitives and the Nvidia OpenShell runtime to enable secure on-device agents. The security layer provides identity, containment, policy, and end-to-end security. OpenShell adds user-defined policies for agent capabilities, intelligent query routing to local vs. cloud models, and PII masking in cloud-bound queries.
Agent frameworks including Hermes Agent and OpenClaw are building Windows apps on this stack, enabling cross-app workflows, file search, image/video generation, and code plugin creation.
Availability
RTX Spark-powered slim laptops (all-day battery, premium displays) and compact desktops will ship this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with Acer and GIGABYTE models to follow.
📖 Read the full source: HN AI Agents
👀 See Also
New AI Tutor Achieves 0.71-1.30 SD Effect Size in Dartmouth Course
A new AI tutor for a Dartmouth introductory CS course showed learning gains of 0.71 to 1.30 standard deviations compared to a control group.
Anthropic Refuses Pentagon Safety Removal Demands, Loses Federal Contracts
Anthropic refused Pentagon demands to remove safety guardrails from Claude for military applications, leading to a $200M contract cancellation and a presidential order banning federal agency use of their technology.
Opus 4.6 excels at research, Gemini 3.1 Pro has better judgment in forecasting benchmark
A benchmark of 1,417 binary forecasting questions separates research and judgment performance: Claude Opus 4.6 leads in agentic research, Gemini 3.1 Pro wins on fixed-evidence calibration. GPT-5.4 and Grok 4.20 show little change between conditions.
Open Source vs Frontier Models: Single-File Canvas Car Scene Benchmark
A developer tested 12 models including GPT-5.5, Claude Opus 4.7, and Qwen 3.6 Plus on a single-file HTML canvas car driving animation task, with results publicly compared.