NTS Elite APEX 2U 4-GPU/ APU MI300A Liquid-Cooled AI Server

Platform overview

The NTS Elite APEX 2U 4-GPU/APU MI300A Liquid-Cooled AI Server (Supermicro AS-2145GH-TNMR-LCC class) packs four AMD Instinct MI300A accelerated processing units into a dense 2U chassis for AI training, LLM inference, generative AI, scientific simulation, climate and weather modeling, and computer-vision pipelines that benefit from CPU-GPU coherence on a shared memory fabric.

Each MI300A combines Zen 4 CPU cores and CDNA 3 GPU compute on one package with onboard HBM3 stacks (512 GB class unified memory across the quad-APU configuration). That design removes a large share of PCIe round-trips that discrete CPU-plus-GPU nodes pay on memory-bound kernels, so architects often choose this SKU when model weights, checkpoint traffic, or tightly coupled HPC solvers dominate wall-clock time.

Thermal headroom comes from direct liquid cooling plus three heavy-duty system fans. NTS requires professional onsite integration with your facility coolant distribution (CDU loops, hoses, and leak detection) so the platform ships as a validated compute node - not a first-time plumbing experiment in the rack.

Key features

• Quad AMD Instinct MI300A APUs with unified HBM3 memory stacks for coherent CPU/GPU workloads
• Optional PCIe layouts up to six PCIe x16 FHHL slots for fabric, storage, or specialty I/O
• Two AIOM networking slots (OCP NIC 3.0 compatible) for Ethernet or InfiniBand adapter choices
• Default eight hot-swap 2.5" NVMe bays, or SAS3/SATA3 bays via storage add-on card; two internal NVMe/SATA M.2 slots
• Liquid-cooled thermal design with three heavy-duty fans; four 2700 W redundant Titanium-level PSUs

Typical workloads

Teams use this node for multi-GPU training jobs that keep tensors close to compute, low-latency inference services that share embedding tables across APU memory, and classical HPC codes (CFD, molecular dynamics, earth-system models) that alternate between host-side orchestration and accelerator kernels. Computer-vision labeling farms and retrieval-augmented generation stacks also fit when PCIe chatter would otherwise starve discrete GPUs.

If your SOW instead calls for discrete Hopper/Blackwell GPUs with separate host CPUs, compare NTS Elite APEX HGX platforms. Choose MI300A when unified memory and liquid-cooled 2U density outweigh discrete GPU SKU flexibility.

Configuration and procurement

Use Configure to select networking, storage media, firmware baselines, and related options. Pricing updates as you change radios and quantities. NTS validates power, cooling, firmware, and drivers in Fremont before ship, and can package the same BOM for SEWP V, ITES-4H, GSA MAS, or SLED cooperatives when your contracting path requires it.

Share target model sizes, fabric preference (Ethernet vs InfiniBand), rack power and coolant budget, and delivery window via contact NTS or Request a Quote. Related reading: AI and GPU infrastructure, data center and HPC, and contract vehicles.