Infinity Forge 4 | High-Performance Multi-GPU Server | 4x RTX PRO 6000 Blackwell Max-Q 96GB Server | EPYC 9575F

For AI teams that have outgrown one or two GPUs. It is designed for organizations serving large open models to many users, labs running four-GPU fine-tuning, render studios and managed service providers hosting private AI for their clients, in a server room or colocation facility with 200 to 240 V power and a rack at least 1,000 mm deep. It delivers four times the GPU memory of a single-GPU Forge 2 in a purpose-built GPU chassis with front-to-back airflow and 1+1 redundant power.

Every Forge 4 is a consultative build, specified around your models, users and site. The reference configuration places four NVIDIA RTX PRO 6000 Blackwell Max-Q 96GB GPUs (384GB of GDDR7 ECC in total) on PCIe 5.0 x16 slots in a 4U single-socket AMD EPYC GPU server platform such as the MSI G4101. It pairs a 64-core AMD EPYC 9575F under a closed-loop liquid cooling module with 384GB of DDR5-6400 ECC registered memory across all twelve channels, dual 25GbE SFP28, two 4TB NVMe drives for the operating system and models, twelve U.2 NVMe bays for expansion, a dedicated BMC port and 1+1 redundant 3000 W power supplies. The platform, GPU count, memory, storage and networking are confirmed with you, and the GPU configuration is confirmed in writing with the platform vendor, before we quote.

Four 96GB GPUs hold Qwen3-235B-A22B-Instruct-2507 in its official FP8 release (236 GB of weights) with a 128K-token context and over 100GB left for concurrent sessions (Infinity estimate), or several large models side by side, such as gpt-oss-120b on one GPU and a 70B model at 8-bit on another. MIG divides the GPUs into as many as sixteen isolated 24GB instances for a shared cluster, and PyTorch FSDP spreads fine-tuning across all four cards.

Compared with Forge 2, it doubles the maximum GPU count to four and moves to a 12-channel EPYC platform with redundant power and 25GbE as standard, but it needs 200 to 240 V power and an 831 mm-deep chassis. Compared with Forge 8, it stays single-socket with 300 W GPUs and draws roughly 2.2 kW, against about 6.9 kW for Forge 8's eight 600 W Server Edition GPUs.

The build includes a processor BIOS check, BMC/IPMI and Redfish setup, a 72-hour burn-in that validates every GPU's PCIe 5.0 x16 link, thermals and power, and the AI stack: NVIDIA open driver and CUDA, NVIDIA Container Toolkit, vLLM configured across the four GPUs and MIG profiles, with an as-built document. A site readiness survey covering voltage, circuits and inlet type, rack depth, cooling and network is required before ordering and is credited against the order.

Options include starting with two GPUs and adding two later; a value and virtual-GPU variant with four RTX PRO 5000 Blackwell 72GB GPUs (288GB, on NVIDIA's vGPU supported list; licences separate); 768GB of system memory where supply allows; dual 100GbE networking; enterprise U.2 NVMe storage; and on-site installation by two technicians with a lift.

It needs 200 to 240 V power on two circuits, a deep rack and a two-person or lift installation, and it has no onboard data network ports, so a network card is always fitted. GPUs connect over PCIe 5.0 without NVLink, Max-Q GPUs do not support NVIDIA vGPU, and a 671B-class model at 4-bit (about 405 GB) does not fit in 384GB. Components are ordered per project and some are non-cancellable, so a deposit is required.

Backed by the Infinity Computers 3-Year Hardware Warranty and Lifetime Technical Support, with no contractual on-site response time. Delivered by pallet freight to a loading dock or server room, or delivered and installed by Infinity by quote. Configured and quoted per project in CAD before HST; built and tested in Ontario.