Supermicro SuperServer SYS-522GA-NRT (RTX PRO 6000/L40s System)

The RTX PRO 6000 Blackwell Server Edition is the current-generation choice, 96GB of GDDR7, FP4 inference, MIG partitioning, and Blackwell throughput. The L40S is the proven Ada-generation universal card at 48GB, right for fleets already standardized on it. Both serve graphics and AI in one card; the decision is generation and memory, and MillionMiner models both in the quote.

High-memory inference at economics no HGX node matches, plus multi-tenant density. Large models serve across cards, long-context work gets real KV cache room, and MIG splits each 96GB card into isolated instances, dozens per chassis. It exceeds the 640GB of an 8x A100 node while the cards also carry the display engines and encoders visual work needs.

It is how the PCIe lanes are wired. Dual-root splits the eight cards across both processors' root complexes, four per CPU, so every GPU gets a direct, uncontended path to host memory. Since this machine's workloads run independently per card and never talk to neighbors, dual-root is the right topology, and PCIe Gen5 doubles the bandwidth on each of those paths.

Primarily visual and inference. Neither the RTX PRO 6000 nor the L40S carries NVLink, so the eight GPUs run independently rather than as a coupled training mesh, which is exactly right for serving, rendering, VDI, and streaming. For distributed training that needs all GPUs synchronized, the catalog's HGX systems are the tool instead.

Both hold eight PCIe GPUs for independent work; the scope differs. The ASUS is the dual-vendor compute-card box, open to NVIDIA and AMD accelerators. This 5U Supermicro is scoped to the universal graphics-and-AI cards, with the thermal volume for eight 600W-class Blackwell boards and a dual-root layout matched to visual fleets. Compute-card inference goes to the ASUS; visual platforms and high-memory RTX inference land here.

The per-card fleet patterns, cloud gaming streamed per card or per partition, VDI carving cards into virtual workstations, render farms per frame, media transcoding per stream, and diagnostic imaging per study. The universal cards carry the display engines, encoders, and driver stacks these need, which is why this class exists apart from compute-only GPUs.

Dual processors with system memory scaling to 6TB, comfortably above the 768GB GPU pool. The live page lists "Intel Xeon or AMD EPYC," which is imprecise for one model, and the exact platform MillionMiner supplies under this SKU is confirmed in the quote along with the memory fit-out.

Data center provisioning for eight 600W-class cards plus a dual-processor host, cooled on air through the 5U chassis whose height exists to move that airflow. The power supply configuration is sized to the card population and confirmed with your quote, and hosting in MillionMiner's facilities is available for teams that prefer not to provision the rack power.

When the workload is distributed training, where all GPUs synchronize and the NVSwitch fabric of the A100, H200, and Blackwell systems in this catalog decides scaling. These universal cards have no NVLink and this platform does not pretend otherwise, it is built for the work where that absence costs nothing. The quote routes the workload honestly.

Submit your workload and target card configuration through the quote form. A MillionMiner specialist confirms the card choice, host, power, and warranty, then delivery. Partial populations grow card by card on the same chassis. Every unit is tested before shipment and delivered worldwide DDP with duties handled.