The ASUS ExpertCenter Pro ET900N G3 redefines what a desktop workstation can achieve for AI development. Unlike traditional GPUs, this system is built around NVIDIA’s Blackwell Ultra Desktop Superchip, which integrates CPU and GPU cores into a single package. The result is a machine that can handle large-scale AI training, inference, and autonomous agent workflows locally—without the need for cloud dependencies.
- Chipset: NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip
- Architecture: NVIDIA DGX Station GB300
- Coherent Memory: 748 GB
- AI Performance: Up to 20 PFLOPS
- Connectivity: NVLink-C2C for high-bandwidth interconnects
The 748 GB of coherent unified memory is a standout feature. This allows developers to work with larger AI models locally, reducing bottlenecks that typically plague conventional workstations. For example, when running the Qwen open-source AI model, the system achieves around 1,600 tokens per second in combined input and output processing—a figure that underscores its capability for demanding workloads.
Why this matters: The ET900N G3 is designed to bridge the gap between enterprise AI needs and on-premises infrastructure. Organizations can now deploy advanced AI capabilities without relying solely on cloud services, offering greater control over data privacy, operational costs, and latency. This is particularly relevant for industries where real-time AI processing is critical.
The system also supports NVIDIA’s NemoClaw workflows, which simplify the development of always-on AI assistants and autonomous agents. Combined with the Blackwell architecture, this positions the ET900N G3 as a future-proof solution for AI research labs, enterprise deployments, and content creation.
Looking ahead: While the ET900N G3 is already available, its long-term impact will depend on software ecosystem growth. NVIDIA’s plans to expand support for Windows-based AI development could further solidify its role in mainstream AI workflows. For now, buyers should focus on whether their workloads require this level of compute power or if cloud alternatives remain more practical.
