.png&w=3840&q=75)
The Power of an AI Supercomputer.
At the Edge.
Powered by the NVIDIA Grace Blackwell architecture. Prototype, run, and fine-tune large models of up to 200B parameters entirely on-premises, without relying on data center or cloud compute constraints.
Monumental Scale.
Minimal Footprint.
Engineered to seamlessly deploy massive parameter models from your local system to distributed edge infrastructure, leveraging up to 1 PetaFLOP of FP4 AI performance.
200B Parameter Support
Execute inference and fine-tuning on massive language models locally, supported by 128 GB of coherent, unified LPDDR5x system memory at 273 GB/s.
ConnectX Scalability
Connect two edge systems via native 200 Gbps ConnectX-7 networking to enable synchronized inference on models up to 405B parameters.
Zero-Setup Software
Preconfigured with NVIDIA NIM microservices, AI Workbench, and the complete CUDA-X library stack out of the box. No dependency tuning required.
Technical Specifications
NVIDIA Grace Blackwell
20 Core Arm (10 Cortex-X925 + 10 Cortex-A725)
1 PetaFLOP (FP4) / 5th Gen Tensor Cores & 4th Gen RT Cores
128 GB LPDDR5x (Coherent Unified Memory)
256-bit Interface / Up to 273 GB/s
4 TB NVMe M.2 with self-encryption
NVIDIA ConnectX-7 NIC (200 Gbps) / 10 GbE RJ-45 / Wi-Fi 7
240 W (Standard Wall Outlet compatible)
150 mm L x 150 mm W x 50.5 mm H / 1.2 kg