Blackwell + Supermicro servers
NVIDIA’s Blackwell family hit the stage with the RTX PRO 4500 Blackwell Server Edition now supported in Supermicro 1U and 2U systems that can hold up to eight GPUs per node for large‑scale inference and media/graphics workloads (investing.com)(prnewswire.com). Supermicro is pitching these nodes for virtualization, AI inference and “AI factory” use cases — ready for datacenter rollouts now (investing.com).
NVIDIA’s RTX PRO 4500 Server Edition ships with 10,496 CUDA cores and 32 GB of GDDR7 ECC memory in a compact single‑slot server form factor. (techpowerup.com) The card’s memory subsystem delivers up to 896 GB/s of bandwidth, supports PCIe Gen5, and exposes fifth‑generation Tensor Cores with FP4 support plus ninth‑generation NVENC and sixth‑generation NVDEC video engines. (nvidia.com) NVIDIA and partners advertise the server edition as a lower‑power option: vendors list a 165 W single‑slot TGP for the server card, while active dual‑slot workstation variants are documented at higher power levels (around 200 W) in third‑party spec sheets. (pny.com) Supermicro’s product announcement on March 16, 2026 names specific platforms that will host Blackwell‑class GPUs — including NVIDIA Vera Rubin NVL72 and HGX Rubin NVL8 architectures — and highlights a 2U Vera CPU design that supports RTX PRO 4500 Server Edition accelerators alongside BlueField‑4 DPU integration. (prnewswire.com) The company said its Data Center Building Block Solutions (DCBBS) liquid‑cooling stack is being engineered for those Rubin platforms and claimed the stack targets up to 10× throughput per watt and one‑tenth the “token cost” versus prior Blackwell configurations. (prnewswire.com) Supermicro positioned the new systems as modular, pre‑engineered rack building blocks for validated NVIDIA reference architectures and called out NVL‑ and HGX‑class designs for scaling inference and long‑context AI workloads. (prnewswire.com)
