ASUS ESC8000A-E12P Dual AMD EPYC 9004 GPU Server

Yes, and that is its defining feature. It supports up to eight dual-slot PCIe GPUs and scales them either way, NVIDIA cards paired over NVLink Bridge or AMD Instinct accelerators over Infinity Fabric Link. It is the only server in this catalog open to both vendors, so your accelerator choice stays a workload decision rather than one your chassis made for you.

Any dual-slot PCIe accelerator, active or passive, up to eight of them, which covers NVIDIA RTX PRO 6000 Blackwell, L40S, and H100 PCIe, along with AMD Instinct cards. The exact models and whether they run active or passive cooling are confirmed with your quote, since the optimal pick depends on your workload.

Two reasons. PCIe lane budget, since eight GPUs plus 13 expansion slots need the lanes two EPYC 9004 sockets provide, and preprocessing, since eight independent GPUs each need data prepared and fed. The 12-channel DDR5 per socket keeps that pipeline from starving the accelerators.

Primarily inference and independent-GPU HPC. The GPUs sit on PCIe rather than a shared NVSwitch baseboard, so they run as eight separate accelerators, ideal for serving, rendering, and parallel HPC where cards do not synchronize. For distributed training that needs all GPUs coupled at fabric speed, the catalog's HGX systems are the right choice instead.

Eight carry the GPUs; the remainder take high-speed fabric NICs, additional storage controllers, or DPUs. That headroom is unusual at this GPU count and is what lets the platform serve as a fleet node with proper networking rather than a GPU box with no room left for the rest of the build.

Yes. Populate the GPU count you need now and grow toward eight as demand rises, adding NVIDIA cards in bridged pairs or Instinct cards as your deployment scales. This pay-as-you-grow path is a core reason to choose a PCIe platform over a fixed eight-GPU baseboard.

They are Tri-Mode, accepting NVMe, SATA, or SAS drives in the same bays, so you can run an all-NVMe dataset tier, mix in SATA capacity, or use SAS, without committing to one interface up front. The fit-out is specified with your quote.

Up to four 3000W Titanium redundant supplies, sized to your GPU selection, since eight accelerators plus a dual-EPYC host is a substantial draw. The exact PSU configuration is confirmed on quote, and MillionMiner advises on the rack power your build requires.

Different architectures for different work. The HGX platforms wire eight GPUs into one NVSwitch mesh for coupled training; this ASUS keeps eight GPUs independent on PCIe for inference and HPC, with the bonus of vendor choice. If your GPUs need to act as one, go HGX; if they each run their own workload, this is the more flexible and economical answer.

Submit your workload and GPU preference through the quote form. A MillionMiner specialist confirms the GPU configuration, host, storage, power, and warranty, then the delivery plan. Every unit is tested before shipment and delivered worldwide DDP with duties and customs handled, with hosting in MillionMiner's own data centers available as the alternative.