Supermicro GPU Systems powered by AMD Instinct™ MI300 Series Accelerators
Supermicro unleashes three new GPU systems powered by AMD Instinct™ MI300 series accelerators to advance large-scale AI and HPC infrastructure. Built on Supermicro’s proven AI building-block system architecture, the new 8U 8-GPU system with MI300X accelerators streamlines deployment at scale for the largest AI models and reduces lead time.
In addition, the 4U and liquid-cooled 2U 4-Way systems supporting MI300A APUs, which combine CPUs and GPUs, leverage Supermicro’s expertise in multiprocessor system architecture and cooling design, finely tuned to tackle the convergence of AI and HPC.
Massive-Scale AI Training and Inference
8U 8-GPU System with AMD Instinct MI300X Accelerators
Fully optimized for the industry-standard OCP Accelerated Module (OAM) form factor, this system provides unparalleled flexibility for rapidly-evolving AI infrastructure requirements and simplifies deployment at scale. A massive pool of 1.5 TB HBM3 per server node erases AI training bottlenecks by containing even extremely-large LLMs within its physical GPU memory, minimizing training time and maximizing the number of concurrent inference instances per node. Designed with full scalability in mind, the system supports 8 high-speed 400G networking cards providing direct connectivity to each GPU for massive AI training clusters.
Converged HPC-AI and Scientific Computing
Air-Cooled 4U Quad-APU System with AMD Instinct MI300A Accelerators
The Supermicro AS -4145GH-TNMR is a 4U 4-way multi-APU system with air cooling for flexible deployment options. It features a balanced CPU-to-GPU ratio for taking on converged HPC-AI applications and is well equipped for a variety of data type precisions. In addition, a total of 912 AMD CDNA™ 3 GPU compute units, 96 “Zen 4” cores, and 512 GB of unified HBM3 memory deliver substantial performance for parallelizable workloads. The mechanical airflow design of this system keeps thermal throttling at bay in addition to 10x heavy-duty 80mm fans.
Enterprise HPC and Supercomputing
Liquid-Cooled 2U Quad-APU System with AMD Instinct MI300A Accelerators
Targeting accelerated HPC workloads, this 2U 4-way multi-APU system with liquid cooling integrates 4 AMD Instinct™ MI300A accelerators. Each APU combines high-performance AMD CPU, GPU and HBM3 memory for a total of 912 AMD CDNA™ 3 GPU compute units and 96 “Zen 4” cores in one system. Supermicro's direct-to-chip custom liquid-cooling technology enables exceptional TCO with over 51% data center energy cost savings. Furthermore, there is a 70% reduction in fans compared to air-cooled solutions. The rack-scale integration optimized with the dual AIOM and 400G networking creates a high-density supercomputing cluster with up to 21 2U systems in a 48U rack.
Proven to Perform at Scale
Industry-Proven System Designs
8U high-performance fabric 8-GPU system leveraging industry-standard OCP Accelerator Module (OAM) to support AMD’s MI300X Accelerators. 4U and 2U multi-processor APU systems integrate 4 AMD Instinct™ MI300A accelerators.
Purpose Built for AI & HPC
Feature maximized and power optimized, supporting up to 16x dedicated hot-swap NVMe drives, full-performance GPUs, CPUs, and memory, and high-speed networking for large-scale cluster deployments.
Advanced Cooling Options
Flexible cooling options for air- and liquid-cooled racks, with liquid-cooled solutions delivering exceptional TCO with over 51% data center energy cost savings.
Designed to Scale
Designed with full scalability in mind, supporting 8 high-speed 400G networking cards providing direct connectivity to each GPU for massive AI clusters.
Optimized Solutions for AI and HPC
Supermicro Systems with AMD Instinct™ MI300 Series Accelerator
 |  |  | |
| Form Factor | 8U 8-GPU System with AMD Instinct MI300X Accelerators (air-cooled) | 2U 4-GPU System with AMD Instinct MI300A Accelerators (liquid-cooled) | 4U 4-GPU System with AMD Instinct MI300A Accelerators (air-cooled) |
| CPU | Dual AMD EPYC 9004 Series Processors with up to 128 cores/256 threads per socket | Quad AMD Instinct MI300A APUs with total of 96 CPU cores (4x 24 AMD “Zen4” cores) | Quad AMD Instinct MI300A APUs with total of 96 CPU cores (4x 24 AMD “Zen4” cores) |
| GPU | 8x AMD Instinct MI300X Accelerators with 192GB HBM3 memory per GPU interconnected on AMD Universal Base Board (UBB 2.0) | 4x AMD Instinct MI300A with 228 AMD CDNA 3 GPU compute units per APU | 4x AMD Instinct MI300A with 228 AMD CDNA 3 GPU compute units per APU |
| Memory | Up to 6TB (24x 256GB DRAM) 4800MT/s ECC DDR5 RDIMM/LRDIMM | 512 GB unified HBM3 memory with up to 5.3 TB/s bandwidth | 512 GB unified HBM3 memory with up to 5.3 TB/s bandwidth |
| Drives | 16x hot-swap PCIe 5.0 U.2 NVMe, 1x onboard M.2 NVMe, 2x 2.5" SATA | 2x onboard 2280 or 22110 M.2 NVMe and 8x hot-swap 2.5" U.2 NVMe | 2x onboard 2280 or 22110 M.2 NVMe and 8x hot-swap 2.5" U.2 NVMe or 24x 2.5" SAS/SATA |
| Networking | 8x PCIe 5.0 high-performance networking cards, up to 400G with Ethernet or InfiniBand | 2x AIOM (OCP 3.0) with up to 400G and additional 4x PCIe 5.0 (x8) slots | 2x AIOM (OCP 3.0) with up to 400G and additional 4x PCIe 5.0 (x8) slots |
| Interconnect | AMD Infinity Fabric™ Links, up to 896GB/s of peer-to-peer inter-GPU connectivity | AMD Infinity Fabric Links, up to 384GB/s of peer-to-peer inter-GPU connectivity | AMD Infinity Fabric Links, up to 384GB/s of peer-to-peer inter-GPU connectivity |
| Power | 6x or 8x 3000W redundant Titanium Level power supplies | 4x 1600W redundant Titanium Level power supplies | 4x 1600W redundant Titanium Level power supplies |
| Cooling | Air Cooling | Liquid Cooling | Air Cooling |