Rack‑level AI infra gets real at NVIDIA GTC
Compal showcased a “One Integrated Solution” rack‑level AI architecture at NVIDIA GTC, signalling a move toward modular, rack‑scale deployments that align with rapid GPU and software cadence for RAG and agentic systems announced. That design favors predictable scaling for enterprise LLM inference and inference‑heavy workflows.
Compal presented a three‑rack physical configuration that separates Compute, Power and Liquid‑Cooling and was exhibited at NVIDIA GTC (Booth 107, March 16–19, 2026) prnewswire.com. The Power Rack in the demo is built on an AcBel‑designed power‑rack architecture, while the Liquid‑Cooling Rack integrates Rayonnant CDU cabinet rack systems and piping infrastructure for coordinated thermal management prnewswire.com. Compal’s compute sleds were shown as high‑density systems based on NVIDIA HGX Rubin NVL8 and NVIDIA HGX B300 platforms — the HGX B300 baseboard is an eight‑Blackwell‑GPU design with up to 288 GB HBM3e per GPU and a total of ~2.30 TB GPU memory per node in reference configurations prnewswire.com. NVIDIA’s HGX B300 reference architecture calls for eight ConnectX‑8 SuperNICs and up to 800 Gbps east–west networking per baseboard, and Compal’s demo pairs those network and baseboard targets with the AcBel power and Rayonnant CDU stacks to illustrate integrated power and thermal management under near‑real‑world data‑center conditions docs.nvidia.com.
