Lenovo ThinkSystem SR675i V3 Dual-Socket 3U AI Inference Rack Server with Dual AMD EPYC Processors, 4 Double-Width GPUs, and 3TB TruDDR5 Memory

The Lenovo ThinkSystem SR675i V3 is a cost-optimized dual-socket 3U rack server tailored for lightweight generative AI inference, computer vision, video analytics, VDI workstation cloud, edge large language model (LLM) serving and offline batch data processing. Built on dual 4th/5th Gen AMD EPYC Genoa/Turin processors, it balances robust multi-thread CPU throughput, flexible mid-tier GPU expansion, ample TruDDR5 memory, low-latency local NVMe storage and energy-efficient air cooling architecture. Unlike high-end liquid-cooled training servers, the SR675i V3 adopts standard air cooling design to cut total deployment cost while supporting up to 4 double-width data center GPUs. Integrated Lenovo XClarity remote management, enterprise-grade RAS reliability and redundant high-efficiency power supplies make it the ideal mainstream AI inference node for smart city visual analysis, enterprise private AI services, retail edge computing and mid-scale AI cluster construction.

Dual AMD EPYC processors deliver abundant CPU cores for data preprocessing, video frame decoding and request scheduling, eliminating CPU bottlenecks for multi-concurrent inference requests. Compared with single-socket GPU servers, the dual-socket design supports more parallel AI tasks, while avoiding the high hardware cost of liquid-cooled full 8-GPU training servers, perfectly matching mid-budget inference business demands.

The SR675i V3 supports four mainstream mid-range data center accelerators, specially tuned for low-latency LLM serving, multi-channel video analytics and VDI rendering workloads. Without expensive NVLink interconnection modules, the platform greatly reduces upfront procurement cost for enterprises that do not require ultra-large unified GPU memory pools for heavy model training.

24 memory slots support up to 3TB high-speed DDR5 memory, allowing users to load multiple lightweight AI models and massive feature datasets entirely inside memory, drastically reducing repeated disk access latency during continuous inference service. Hardware ECC and memory sparing prevent service crashes caused by transient memory errors for 7×24 unattended cluster operation.

10 front hot-swap U.3 NVMe bays provide local high-speed flash storage to cache frequently accessed AI training datasets, inference media files and model weights. Universal AnyBay trays support flexible capacity expansion without server power-off, simplifying daily dataset update and storage maintenance for visual AI and retail edge scenarios.

This model cancels complex liquid cooling pipelines and adopts optimized high-static-pressure air cooling fans, cutting hardware procurement and later maintenance costs. The refined airflow design effectively dissipates heat generated by 4 mid-power GPUs under sustained inference load, meeting the stable heat dissipation needs of general data center environments without dedicated liquid cooling infrastructure.

Modular Platinum redundant power supplies support N+1 fault-tolerant configuration, avoiding unexpected downtime of online AI services caused by single power supply damage. Ultra-high energy efficiency reduces long-term power consumption for large-scale deployed inference clusters and lowers overall data center operating expenses.

The built-in XCC chip continuously tracks GPU running status, utilization and temperature, sending instant alarms for hardware abnormalities to prevent inference service interruption. XClarity Administrator realizes one-click batch deployment, firmware upgrade and resource scheduling for hundreds of SR675i V3 inference nodes, greatly reducing manual operation workload for IT teams managing multi-node AI platforms.

• 3U AI Inference Rack Server: Cost-balanced multi-GPU chassis optimized for online model serving rather than heavy training
• Air-Cooled Mid-GPU Architecture: Standard fan cooling design without liquid cooling to lower total cost of ownership
• TruDDR5 3DS RDIMM: Lenovo certified stacked high-capacity enterprise memory with built-in error correction
• AnyBay Hot-Swap Tray: Universal drive carrier compatible with NVMe, SAS and SATA storage media
• OCP 3.0 Mezzanine Slot: Modular expansion port for high-speed cluster networking adapters
• CRPS Platinum PSU: Hot-swappable high-efficiency redundant power supply supporting N+1 fault tolerance
• XCC (XClarity Controller): Lenovo exclusive out-of-band server hardware monitoring and remote control system

1. White background full front chassis render (Homepage top banner main image)
2. Front U.3 NVMe drive bay close-up (Dataset caching & storage specification section)
3. Internal GPU & CPU layout diagram (4-GPU inference acceleration selling point block)

The Lenovo ThinkSystem SR675i V3 dual-socket 3U air-cooled AI inference server delivers cost-effective customizable configurations for mid-tier AI workloads. Entry barebone units (dual low-core EPYC, 256GB DDR5, no GPUs, single Platinum PSU, 3-year warranty) start at USD $7,900 / CNY ¥56,000. Standard mainstream inference nodes equipped with dual mid-range AMD EPYC CPUs, 1TB DDR5 memory, NVMe flash storage array and 2× L40S GPUs range from USD $28,500–$52,000. Fully expanded high-concurrency inference servers with top-tier dual EPYC processors, maximum 3TB memory, full 10-bay NVMe storage, 4× double-width GPUs and N+1 redundant Platinum power supplies can exceed USD $86,000. Final pricing is determined by CPU tier, memory capacity, GPU model & quantity, storage layout, power supply redundancy mode, high-speed network add-ons, warranty service tiers and regional enterprise bulk order discounts.