Skip to main content

NVIDIA RTX Spark (AI PC Superchip)

Product Overview​

NVIDIA RTX Spark was officially unveiled on June 1, 2026 at Computex 2026 / GTC Taipei as NVIDIA's first AI PC superchip developed in collaboration with MediaTek. It adopts a unified-memory Arm CPU + Blackwell GPU architecture with 128GB LPDDR5X of shared memory, aiming to turn Windows into an "Agentic AI" platform.

RTX Spark targets both laptops and compact desktops and is expected to launch in Fall 2026, with launch OEMs including Dell, HP, Lenovo, Asus, MSI, and Microsoft Surface.

Core Specifications​

ParameterValue
ArchitectureBlackwell GPU + Arm CPU (MediaTek collaboration)
CPUUp to 20 Arm cores
GPU6,144 CUDA cores (Blackwell)
Unified Memory128 GB LPDDR5X (shared by CPU + GPU)
Memory Bandwidth300 GB/s
InterconnectNVLink C2C (CPU-GPU)
Gaming Performance~100 FPS @ 1440p (DLSS 4.5)
Model CapacityRuns 120 billion parameter models
ContextUp to 1 million tokens
AI Compute (estimated)~1 PFLOPS (unofficial NVIDIA figure)
TDP~100W (estimated)
Launch OEMsDell, HP, Lenovo, Asus, MSI, Microsoft Surface
Models30+ laptops + ~10 desktops
AvailabilityFall 2026
PricingNot yet announced
Announced2025-01-07 (Project DIGITS)

⚠️ Note: NVIDIA's Computex 2026 announcement focused on platform positioning and the AI narrative; it did not publish official TFLOPS/PFLOPS figures, TDP, or pricing. Some of the data above is based on industry estimates.

Platform Ecosystem​

PartnerCollaboration
MicrosoftOpenShell framework, turning Windows into an AI Agent platform
AdobePhotoshop / Premiere core rewrites, 100% GPU-accelerated
MediaTekCo-design of the Arm CPU architecture
Dell, HP, Lenovo, Asus, MSILaunch OEM laptops/desktops

Comparison with DGX Station for Windows​

MetricRTX SparkDGX Station for Windows
ChipRTX Spark SuperchipGB300 Grace Blackwell Ultra
Memory128GB LPDDR5X748GB (252GB HBM3e + 496GB LPDDR5X)
Memory Bandwidth300 GB/s7.1 TB/s (GPU)
CUDA Cores6,144~20,000 (estimated)
AI Compute~1 PFLOPS20 PFLOPS FP4
TargetPersonal AI agents / laptopsDesktop AI supercomputing / trillion-parameter models
TDPTBD1,600W
AvailabilityFall 20262026 Q4

Vendor Information​

ParameterDetails
ManufacturerNVIDIA × MediaTek
Official Websitehttps://www.nvidia.com
AnnouncementComputex 2026 / GTC Taipei (2026-06-01)
AvailabilityFall 2026

Use Cases​

  • ✅ Personal AI agents: run 120B parameter models locally with 1 million token context
  • ✅ Creator workstations: Adobe Photoshop / Premiere 100% GPU-accelerated
  • ✅ Local LLM inference: 128GB unified memory suffices for most models
  • ✅ Light AI training: fine-tuning / LoRA / small-scale training
  • ❌ Large-scale training (DGX Station is a better fit)
  • ❌ HPC scientific computing (H100 / Rubin are better fits)