Lenovo ThinkSystem SR685a V3 Server

CPU: Dual EPYC 9005 or 9004

GPU:  8x Eight AMD MI300X 750W OAM GPUs with 192GB HBM3 GPU memory per GPU

MEM: Up to 3TB with 24x 128GB RDIMMs or 3DS RDIMMs

STO: 245.76TB using 16x 15.36TB 2.5-inch NVMe SSDs

EXP: 10x PCIe 5.0 x16 slots

PWR:  Up to eight hot-swap redundant AC power supplies with up to N+N redundancy. 80 PLUS Titanium certification. 2600 W AC power supplies requiring 220 V AC supply.

Download Data Sheet

Example Use Cases:

• Large language model training and fine-tuning
  The platform supports up to eight AMD MI300X, NVIDIA H100, or NVIDIA H200 GPUs with very high HBM memory capacity and high-speed GPU-to-GPU interconnects, which are exactly the traits needed for distributed LLM training and parameter-heavy fine-tuning.

• Generative AI inference for large or long-context models
  The SR685a V3 is well suited for production inference of large foundation models because of its eight-GPU design, large aggregate GPU memory, fast NVLink/Infinity Fabric connectivity, and high-bandwidth networking for clustered inference deployments.

• HPC simulation and scientific computing
  Lenovo directly calls out simulation and scientific research, and the system’s dual EPYC CPUs, large memory footprint, GPU acceleration, PCIe 5.0 I/O, and 100/200/400Gb networking make it a strong fit for physics, engineering, computational chemistry, and similar HPC workloads.

• Computer vision and multimodal AI training
  Vision, video, and multimodal training pipelines benefit from the server’s dense GPU configuration, large HBM capacity, fast local NVMe storage, and high-throughput data movement between storage, CPUs, GPUs, and network fabric.

• Financial modeling, risk analytics, and accelerated data science
  Lenovo specifically lists financial tech among suggested uses, and the combination of many-core EPYC processors, large DDR5 memory, multiple top-end GPUs, and fast networking makes it a strong platform for Monte Carlo simulation, portfolio/risk analytics, and large-scale model training in quantitative environments.

Bottom line