Lenovo ThinkSystem SR675 V3 3U AI GPU Rack Server with 8 High-End GPU Accelerators, 3TB TruDDR5 Memory, and Neptune Liquid Cooling

The Lenovo ThinkSystem SR675 V3 is a powerful 3U rack GPU server engineered exclusively for AI training, large model inference, high-performance computing (HPC), 3D rendering, computer vision and generative AI workloads. Powered by dual 4th/5th Gen AMD EPYC Genoa/Turin processors, it delivers massive multi-threaded CPU performance paired with support for up to 8 high-end double-wide PCIe/SXM5 GPU accelerators, full PCIe 5.0 I/O, ultra-large TruDDR5 memory capacity and flexible local NVMe flash storage. Equipped with Lenovo Neptune liquid cooling options, high-bandwidth NVLink GPU interconnect, redundant titanium power supplies and unified XClarity remote management, this 3U dense AI node cuts deployment footprint and total cost of ownership for AI labs, cloud AI service providers, research institutions and enterprise big data centers.

Full Technical Specification Table

Core Product Highlights

1. Dual AMD EPYC High-Core Compute Foundation
Dual 5th Gen AMD EPYC Turin CPUs deliver up to 320 concurrent threads, eliminating CPU bottlenecks for large dataset preprocessing, data loading and AI model training. The high PCIe 5.0 lane count fully unlocks GPU accelerator throughput without I/O bandwidth limitations, far exceeding single-socket GPU servers for heavy batch AI workloads.

2. Up to 8 High-End GPU Accelerators with NVLink
The SR675 V3’s 3U dense design supports up to eight double-wide datacenter GPUs, creating a compact high-performance AI node. Built-in NVLink high-speed interconnection enables combined multi-GPU memory pools, drastically accelerating large language model (LLM) training, image generation and 4K/8K video rendering tasks compared to standard PCIe-only GPU servers.

3. Up to 3TB TruDDR5 High-Speed Memory
24 DIMM slots support a maximum of 3TB DDR5 memory at up to 6400MHz, ideal for loading ultra-large training datasets and massive embedding models entirely in memory. Built-in ECC protection prevents training interruption caused by memory bit errors, ensuring stable long-duration AI training jobs.

4. Optional Lenovo Neptune Liquid Cooling for Full GPU Performance
The Neptune hybrid liquid cooling variant eliminates thermal throttling under sustained full GPU load. It maintains lower rack PUE, reduces data center cooling energy costs, and allows GPUs to run at maximum rated clock speeds non-stop for continuous training workloads, a critical advantage for 24/7 AI cluster operation.

5. Flexible High-Speed Local NVMe Storage
10-bay front U.3 NVMe flash storage delivers microsecond-level low latency for local dataset caching, reducing repeated remote storage access delays during model training. Universal AnyBay trays support mixed SSD deployment for flexible storage capacity expansion without server power-off.

6. Titanium High-Output Redundant Power Supplies

High-wattage 2400W Titanium redundant PSUs provide stable continuous power for full GPU and CPU load operation. Hot-swappable modular design avoids single power source failure risks, safeguarding long-running AI training tasks from unexpected downtime.

7. Intelligent XClarity AI Cluster Lifecycle Management

The XCC remote management chip tracks GPU temperature, power draw and utilization metrics in real time, sending instant alerts for overheating or hardware faults. XClarity Administrator supports mass deployment, firmware patching and resource scheduling for hundreds of SR675 V3 AI nodes, cutting IT operation overhead for large-scale AI clusters.

Terminology Glossary

• 3U AI GPU Rack: Compact 3-unit rack chassis optimized for multi-GPU accelerator deployment
• Neptune Liquid Cooling: Lenovo hybrid liquid cooling system for high-power GPU thermal control
• NVLink: High-bandwidth direct GPU-to-GPU interconnect for unified memory acceleration
• TruDDR5 3DS RDIMM: Lenovo certified stacked high-capacity enterprise memory with ECC correction
• OCP 3.0: Modular mezzanine slot for high-speed 100GbE AI cluster networking adapters
• CRPS Titanium PSU: High-efficiency hot-swappable redundant power supply for heavy GPU power demand
• XCC (XClarity Controller): Lenovo exclusive out-of-band server hardware monitoring and management system

3 HD Watermark-Free Official Images for Product Page Layout

1. White background full front chassis render (Homepage top banner main image)
2. Front NVMe drive bay close-up (Storage & dataset caching feature section)
3. Internal GPU array layout diagram (AI acceleration & NVLink selling point block)

Page Layout Matching Guide

1. Top Banner: Place the white-background full chassis render with full product title
2. Storage Specification Table: Insert drive bay close-up beside parameter table
3. Core AI GPU Performance Section: Display internal GPU layout diagram below feature text

Concise Global Price Summary for Listing

Price Overview

The Lenovo ThinkSystem SR675 V3 dual-socket 3U AI GPU server supports fully customized configurations with a wide price spectrum. Entry air-cooled barebone nodes (dual low-core EPYC, 256GB DDR5, no GPUs, single Platinum PSU) start at USD $9,600 / CNY ¥68,000. Standard mid-tier AI servers with dual mid-range EPYC CPUs, 1TB memory, local NVMe flash storage and 4×L40S GPUs range from USD $38,000–$75,000. Fully maxed liquid-cooled training nodes with top-tier dual EPYC processors, 3TB full memory, 8×NVIDIA H200 GPUs, NVLink interconnection and 2400W Titanium redundant power supplies can exceed USD $180,000. Final pricing depends on CPU generation, memory volume, cooling type, GPU model & quantity, storage layout, power supply selection, high-speed network add-ons, and regional bulk channel discounts.