Aiserver Manufacturer
1 / 5
| System Architecture |
• Chipset: GB10 Grace Blackwell Superchip • Combines a 20-core Arm CPU with a Blackwell GPU via NVLink-C2C interconnect • Coherent unified memory architecture for CPU-GPU data sharing |
| CPU |
20-core Arm v9 processor • 10× Cortex-X925 cores (high-performance) • 10× Cortex-A725 cores (energy-efficient) |
| GPU |
Blackwell architecture with: • 6,144 CUDA cores (equivalent to RTX 5070)710 • 5th-generation Tensor Cores (1,000 AI TOPS FP4 precision) • 4th-generation RT Cores (ray tracing acceleration) |
| Video Engines | 1× NVENC (9th gen), 1× NVDEC (5th gen) |
| System Memory |
28GB LPDDR5X unified memory • 256-bit interface, 273GB/s bandwidth • Supports models up to 200B parameters locally, 405B parameters with dual-node clustering |
| Storage | 1TB or 4TB NVMe M.2 SSD (self-encrypting) |
| Networking |
• ConnectX-7 Smart NIC (200Gbps RDMA for clustering) • 10GbE RJ45 port • Wi-Fi 7 (802.11be) + Bluetooth 5.3 |
| Interfaces |
• 4× USB4 Type-C (40Gbps) • 1× HDMI 2.1a (supports 8K video output) • Audio: HDMI multichannel audio |
| Power Consumption |
• 170W typical operation916 • Up to 224W peak under heavy load7 |
| Cooling | Passive airflow design optimized for silent operation |
| Physical Specifications |
• Dimensions: 150mm (L) × 150mm (W) × 50.5mm (H) • Weight: 1.2kg • Form Factor: FHFL dual-slot (rack-mountable with optional kit) |
| Operating System |
DGX OS (Ubuntu 22.04-based) • Pre-installed AI Enterprise software stack • Supports frameworks: PyTorch, TensorFlow, RAPIDS |
| Model Capabilities |
• Inference: Up to 200B parameters (e.g., LLaMA 2, DeepSeek-R1) • Fine-tuning: Up to 70B parameters • Distributed training: Scale to 405B parameters with dual DGX Spark nodes116 |
| Security |
• Secure Boot and ECC memory protection • Confidential computing with ARM TrustZone |
| Development Tools |
• AI Enterprise Suite (cuDNN, TensorRT, etc.) • NGC container registry access |
