ASUS ExpertCenter Pro ET900N G3 Brings 748GB of Coherent Memory to Desk-Side AI

Editorial illustration of a large silver AI workstation with visible cooling hardware and streams of light converging on a shared memory core.

ASUS has released the ExpertCenter Pro ET900N G3, a desk-side AI workstation built around Nvidia's GB300 Grace Blackwell Ultra Superchip. Its defining feature is 748GB of coherent memory shared across the CPU and GPU, giving large models and many concurrent AI agents access to a common memory pool without relying on a conventional split between system and graphics memory.

Memory Takes Priority

The workstation combines 72 Arm CPU cores with a Blackwell Ultra GPU in one chassis. In testing reported by TechRadar, a GLM-5.2 workload placed 215GB of expert weights in the coherent memory fabric, illustrating the system's intended role: running models whose memory requirements and data movement can constrain standard workstations.

The same architecture is designed for highly concurrent inference. The reported results included 4,793 output tokens per second across 128 streams in a MiniMax M2.7 benchmark. A separate GPT-OSS-120B demonstration reached 9,280 output tokens per second and more than 45,000 tokens per second in total throughput. These figures come from the cited reviewer demonstrations rather than independent Signal Titans testing.

Two Systems Can Pool Memory

Two 400-gigabit networking ports provide up to 800Gb/s of combined bandwidth for a direct connection to a second GB300 station. According to the report, a linked pair can expose a 1,496GB coherent memory pool, extending the approach to models that exceed the capacity of one machine.

Built for a Fixed Location

This is not a conventional desktop tower. The ExpertCenter Pro ET900N G3 weighs 27kg and measures 584 x 232 x 565 mm. ASUS equips it with a sealed liquid-cooling system, copper cold plates and dedicated airflow for the network card's optical transceivers. Power comes from a 1,600-watt 80 PLUS Titanium supply, while one demonstration reportedly approached 1,000 watts under a workload involving dozens of agents.

The storage configuration uses two mirrored 2TB NVMe drives, with three PCIe slots and two additional expansion bays available for upgrades. Two metal handles help staff move the heavy chassis, although its dimensions, cooling requirements and power draw make it better suited to a permanent desk-side installation.

Pricing Remains an Estimate

ASUS has not disclosed an official price. Estimates cited by TechRadar place the workstation between $100,000 and $120,000, limiting its likely audience to organizations that need local access to very large models or dense multi-agent workloads and can support the accompanying power, cooling and infrastructure demands.

AI-generated editorial illustration; not a documentary photograph.

Source: An Asus homage to legendary Apple Mac Pro: GB300 DGX workstation gets two thumbs up from reviewer who calls $120,000 AI supercomputer 'the most capable thing we’ve ever put on a desk'

Share this article
FacebookThreadsLinkedInX

Leave a Reply

Your email address will not be published. Required fields are marked *