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
| Parameter | Value |
|---|
| Architecture | Blackwell GPU + Arm CPU (MediaTek collaboration) |
| CPU | Up to 20 Arm cores |
| GPU | 6,144 CUDA cores (Blackwell) |
| Unified Memory | 128 GB LPDDR5X (shared by CPU + GPU) |
| Memory Bandwidth | 300 GB/s |
| Interconnect | NVLink C2C (CPU-GPU) |
| Gaming Performance | ~100 FPS @ 1440p (DLSS 4.5) |
| Model Capacity | Runs 120 billion parameter models |
| Context | Up to 1 million tokens |
| AI Compute | 1 PFLOPS (FP4 sparse, official figure) |
| TDP | ~100W (estimated) |
| Launch OEMs | Dell, HP, Lenovo, Asus, Acer, Gigabyte, Microsoft, MSI (8 OEMs) |
| Models | 30+ laptops + ~10 desktops |
| Availability | Fall 2026 |
| Pricing | Not yet announced |
| Announced | 2025-01-07 (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.
DGX Spark Desktop System (October 2026 Update)
The first official desktop form factor of the RTX Spark platform, DGX Spark 64GB, launched in early October 2026:
| Item | Spec |
|---|
| Chip | GB10 Grace Blackwell superchip (fully integrated) |
| Unified memory | 64 GB (~8 GB reserved for system, ~56 GB available for models) |
| Memory bandwidth | 273 GB/s |
| Interconnect | NVLink-C2C |
| Local inference | Runs 100B-parameter models locally (Qwen3.8 27B, Gemma4 26B, Nemotron 3.5 Lightning verified) |
| Dual-node clustering | Built-in ConnectX-7; two units linked via QSFP pool memory to 128 GB, raising the model limit to ~200B parameters |
| Dual-node benchmark | Qwen3.8 27B cluster performance up to 70% higher than a single 128GB system peak |
| Starting price | $4,999 |
| Availability | October 23, 2026 (Acer, ASUS, Dell, Gigabyte, HP, MSI and more) |
| 128GB Founders Edition | Price raised to $6,950 (AI memory shortage driving hardware pricing) |
| Software | Ships with DGX OS + NVIDIA AI stack; native Ollama, vLLM, PyTorch with CUDA support; Sync Cluster Assistant for automated cluster setup |
| Partner | Collaboration |
|---|
| Microsoft | OpenShell framework, turning Windows into an AI Agent platform |
| Adobe | Photoshop / Premiere core rewrites, 100% GPU-accelerated |
| MediaTek | Co-design of the Arm CPU architecture |
| Dell, HP, Lenovo, Asus, MSI | Launch OEM laptops/desktops |
Comparison with DGX Station for Windows
| Metric | RTX Spark | DGX Station for Windows |
|---|
| Chip | RTX Spark Superchip | GB300 Grace Blackwell Ultra |
| Memory | 128GB LPDDR5X | 748GB (252GB HBM3e + 496GB LPDDR5X) |
| Memory Bandwidth | 300 GB/s | 7.1 TB/s (GPU) |
| CUDA Cores | 6,144 | ~20,000 (estimated) |
| AI Compute | ~1 PFLOPS | 20 PFLOPS FP4 |
| Target | Personal AI agents / laptops | Desktop AI supercomputing / trillion-parameter models |
| TDP | TBD | 1,600W |
| Availability | Fall 2026 | 2026 Q4 |
| Parameter | Details |
|---|
| Manufacturer | NVIDIA × MediaTek |
| Official Website | https://www.nvidia.com |
| Announcement | Computex 2026 / GTC Taipei (2026-06-01) |
| Availability | Fall 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)