NVIDIA
ConnectX-8 SuperNIC
Up to 800Gb/s for multi-system scale-out
Built from the ground up to build and run AI, the ET900N G3 unleashes the NVIDIA® GB300 Grace Blackwell Ultra Desktop Superchip and 748GB of large coherent memory - enough to prototype, fine-tune and infer models of up to 1 trillion parameters at the deskside.
Secure, production-ready agentic Al out of the box
Links two systems to handle even larger models
Preconfigured stack, ready on first boot
Grace 72-Core Neoverse V2 processor · up to 4 x M.2 2280 NVMe PCle 5.0 SSD, 8TB total
Keep regulated and proprietary data inside your own four walls. The ET90ON G3 runs training and inference on-premises, with predictable cost and full control over where your data lives.
Start at a single deskside system and grow into a cluster. QSFP 400G networking and the NVIDIA ConnectX-8 SuperNIC move data between systems fast enough to train on models no single machine could hold.
A data center-grade thermal design pulls cool air across the superchip and exhausts it clear of the chassis, so sustained AI workloads hold their clocks.
Inventory, monitoring, deployment and patching from one console
Preconfigured and supportable on day one
Out-of-band management over 1G LAN, Micro-B or MiniDP
Business-class warranty and service network
Prototype, fine-tune and run fleets of AI agents on your own hardware - without sending a single token off-site.
Secure, production-ready agentic AI, preconfigured
Holds whole models and long agent context
1Power button
2Message LEDs
3Audio
4USB 10G Type-C
5USB 10G Type-A
6USB 2.0
1QSFP 400G ports
2USB 10G Type-A
31G LAN for BMC
410G LAN
5Audio
6Micro-B port for BMC
7MiniDP for BMC





















