Flexible 4U AMD EPYC GPU Platform — Up to 8 PCIe GPUs for AI, HPC, and Mixed Workloads
The AS-4125GS-TNRT is Supermicro's versatile 4U AMD EPYC GPU SuperServer supporting up to 8 double-width PCIe GPUs across a wide compatibility matrix including NVIDIA H100, H100 NVL, L40, L40S, A100, A30, RTX 6000 Ada, and AMD Instinct MI210. Dual AMD EPYC 9005/9004 Series CPUs with PCIe 5.0, up to 6TB DDR5, and 4× 2,000W Titanium power supplies make it an adaptable, multi-vendor GPU platform for organisations requiring flexibility across GPU generations.
Get Pricing — Speak to Our Team
Competitive pricing · Response within 4 hours
Unlike HGX-based systems locked to NVIDIA SXM GPUs, the AS-4125GS-TNRT supports 8 PCIe-slot GPUs from multiple vendors: NVIDIA H100 PCIe (80GB), H100 NVL (96GB, dual-die), L40S (48GB, ideal for inference), RTX 6000 Ada (48GB), A100 (80GB), A30 (24GB), A16, as well as AMD Instinct MI210 (64GB). This multi-vendor flexibility protects against supply chain constraints and allows GPU selection to match specific workload requirements.
AMD EPYC 9005 Series (Zen 5, up to 192 cores per socket) and EPYC 9004 Series (Zen 4, up to 96 cores per socket) deliver exceptional CPU-side compute alongside GPU acceleration. With up to 160 total system cores (dual EPYC 9965X) and 512MB of L3 cache per socket, the host CPU can perform substantial data preprocessing, model compilation, and inference post-processing without creating CPU bottlenecks for GPU-bound workloads.
PCIe 5.0 x16 slots deliver 128 GB/s bidirectional bandwidth per GPU — double the bandwidth of PCIe 4.0. This is critical for workloads that frequently transfer large tensors between CPU DRAM and GPU HBM, including reinforcement learning from human feedback (RLHF), model inference with dynamic batching, and HPC pre/post-processing pipelines that interleave CPU and GPU computation.
24 DIMM slots supporting up to 6TB of DDR5 ECC memory (6000MT/s with EPYC 9005) provide exceptional host memory bandwidth and capacity. This enables in-memory preprocessing of large training datasets, large CPU-side caches for GPU inference serving, and memory-intensive HPC workloads that exceed typical GPU HBM capacities and require CPU-side buffering.
For NVIDIA H100 PCIe and selected GPUs, optional NVLink Bridges connect adjacent GPU pairs with dedicated high-bandwidth interconnect — enabling tensor parallelism between paired GPUs without consuming PCIe bandwidth. AMD Infinity Fabric™ Link is optionally available for AMD Instinct GPU pairs. This makes the AS-4125GS-TNRT suitable for both single-GPU workloads and multi-GPU tensor-parallel inference serving.
All GPU options in the AS-4125GS-TNRT are air-cooled via PCIe card blowers and chassis fans. The system operates in standard data centre air-cooling environments without CDU or liquid cooling modifications — making it suitable for colocation deployments, retrofitting into existing facilities, and use cases where liquid cooling infrastructure investment is not justified by workload requirements.
The Supermicro AS-4125GS-TNRT is a 4U dual-socket AMD EPYC GPU SuperServer designed for workloads requiring flexible GPU selection, high CPU core counts, and large host memory capacity alongside multi-GPU acceleration. It supports up to 8 double-width, full-length PCIe GPU cards with a compatibility matrix spanning NVIDIA H100, L40S, A100, RTX series, and AMD Instinct MI210.
The platform's flexibility stems from its PCIe-native GPU architecture — all GPUs connect via standard PCIe 5.0 x16 slots rather than a proprietary SXM/HGX baseboard. While this means GPU-to-GPU communication uses PCIe bandwidth rather than NVLink at full NVSwitch speed (except where NVLink Bridges optionally connect pairs), it enables multi-vendor GPU support and future GPU upgradability without chassis replacement.
Dual AMD EPYC processors (up to 192 cores per socket with 9005 Series, or 96 cores per socket with 9004 Series) give the AS-4125GS-TNRT exceptional CPU-side processing capability relative to Intel Xeon Scalable alternatives. This is particularly beneficial for workloads with significant CPU pre/post-processing: reinforcement learning environments, simulation-based training, data-intensive inference pipelines, and HPC codes with balanced CPU/GPU parallelism.
Storage flexibility includes 2 hot-swap 2.5" SATA bays, 4 hot-swap 2.5" NVMe bays, and 1 M.2 slot for OS boot. Onboard 2× 10GbE provides baseline connectivity; the available PCIe 5.0 slots support 400GbE NIC cards for high-bandwidth cluster networking. Power is delivered by 4× 2,000W Titanium-level supplies (2+2 or 3+1 redundancy).
Organisations running diverse inference workloads benefit from the AS-4125GS-TNRT's multi-GPU compatibility. L40S GPUs (48GB, optimised for inference throughput) can be deployed for high-concurrency LLM serving; H100 PCIe cards for latency-sensitive applications; A30 or A16 for lighter models or virtual GPU (vGPU) deployments. A single chassis can be configured for the specific GPU mix that matches the inference portfolio.
AMD EPYC's memory bandwidth advantage (up to 460 GB/s per socket with EPYC 9005 and 12-channel DDR5) over Intel Xeon makes the AS-4125GS-TNRT particularly attractive for HPC codes that require strong CPU-side memory performance alongside GPU acceleration — computational fluid dynamics, molecular dynamics, seismic processing, and finite element analysis all benefit from balanced CPU+GPU bandwidth.
Research environments that iterate across multiple GPU generations and experiment with different GPU types — comparing H100 and L40S for specific inference tasks, testing AMD MI210 for compatibility, or evaluating different GPU memory configurations — benefit from the AS-4125GS-TNRT's flexibility. The ability to reconfigure GPU hardware without changing the chassis makes it a sustainable research platform across hardware generations.
Trusted UK IT reseller since 2003 — supplying businesses of all sizes with genuine, competitively priced technology.
Servnet is an authorised IT reseller with direct access to leading product lines at competitive pricing.
Our engineers will help you select the right product and advise on compatibility with your existing infrastructure.
Most products ship from UK stock. Express delivery available for urgent deployments.
All products are genuine, brand-new, with full manufacturer warranty and RMA support.
We can pre-configure your hardware before shipping — reducing on-site deployment time and cost.
Post-sale support from our certified engineers. We're here if anything goes wrong.
