Infinity Forge 8 | Enterprise AI Research Server | 8x RTX PRO 6000 Blackwell Server Edition 96GB Server | 2x EPYC 9575F

For university and hospital research groups, AI start-ups and enterprises that need sustained eight-GPU capacity with data kept on site: an on-premises alternative to shared HPC allocations or long-term cloud GPU rental. It is a data-centre system that draws about 6.9 kW, needs four 220 to 240 V inputs and occupies 4U of a deep rack, so every project begins with a consultation and a site survey.

The reference design is built on the ASUS ESC8000A-E13, a 4U dual-socket GPU server in which each of eight dual-slot GPU slots has its own PCIe 5.0 x16 link to a processor. It pairs two 64-core AMD EPYC 9575F processors and 768GB of DDR5-6400 ECC registered memory with eight NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs (96GB of GDDR7 ECC each, 768GB in total), cooled by the chassis's GPU fan modules and configurable up to 600 W each. Dual 100GbE networking, two 3.84TB enterprise U.2 NVMe drives and 3+1 redundant 3200 W Titanium power supplies complete it. The platform, GPU model and count, memory and fabric are confirmed for each project.

768GB of GPU memory holds a 671B-parameter open model such as DeepSeek-V3.1 at 4-bit (Q4_K_M, 405 GB) with ample room for long contexts and many users (Infinity estimate), or several large models in parallel. The Server Edition supports NVIDIA vGPU, so the node can host GPU-accelerated virtual machines and desktops, and MIG can create up to 32 isolated 24GB instances for a shared research cluster.

Compared with Forge 4, it doubles the GPU count and moves to 600 W Server Edition GPUs that support NVIDIA vGPU, two processors with 24 memory channels, 3+1 redundant power and 100GbE. It is not an NVSwitch training system: GPUs communicate over PCIe 5.0, which suits inference, fine-tuning and many parallel jobs.

Each project includes assembly, BIOS and BMC setup, GPU power-cap configuration matched to your power feeds, a 72-hour burn-in validating every GPU's link, thermals and power, the CUDA stack with the NVIDIA Container Toolkit, optional vGPU host setup and an as-built document. A site survey covering four 220 to 240 V / 16 A inputs, PDU sizing, about 23,500 BTU/h of cooling, rack depth and floor loading is mandatory; on-site installation, Slurm or Kubernetes setup and a remote-monitoring arrangement are quoted.

Options include starting with four GPUs and expanding to eight, eight RTX PRO 5000 Blackwell 72GB GPUs (576GB) subject to platform-vendor confirmation, 1.5TB of system memory where supply allows, a RAID/HBA card for more NVMe drives, and 400 Gb fabric for multi-node work.

It is for data centres only: 220 to 240 V power, about 6.9 kW and about 23,500 BTU/h of heat, a deep rack and a lift installation. The 3+1 power supplies protect against a supply failure; surviving the loss of a whole A/B power feed at full load needs a GPU power cap or more feed capacity. There are no onboard data ports and no NVLink, and only two NVMe drives without a RAID/HBA card. Platform and Server Edition GPU availability, pricing and lead time are confirmed at quote; there is no stock unit.

Backed by the Infinity Computers 3-Year Hardware Warranty and Lifetime Technical Support, with no contractual on-site response time. Delivered by pallet freight with liftgate and inside delivery to the data centre, or installed by Infinity by quote. Every order is quoted in CAD before HST, with a deposit and agreed terms; built and tested in Ontario.