Lenovo ThinkSystem SC777 V4 Neptune Server: Early Performance Overview

From scientific simulation and engineering analysis to AI model training and inference, organizations increasingly require a common platform capable of supporting diverse workloads. The Lenovo ThinkSystem SC777 V4 brings together the technologies needed to accelerate both HPC and AI on a single system. The article will examine the platform design, discuss the types of HPC, AI, and AI-for-Science workloads it is designed to support, and present initial functional and application-level performance results.

The results presented in this paper provide an early look at performance on the SC777 V4 platform. Testing was conducted using pre-production laboratory systems and software environments available at the time of evaluation as the platform approaches general availability. As with any new platform, performance is expected to continue evolving as hardware, firmware, drivers, libraries, and applications mature. The results should therefore be viewed as directional indicators that highlight platform capabilities and workload behavior rather than final optimized performance levels.

The Lenovo ThinkSystem SC777 V4 is designed from the ground up as a next-generation hybrid platform for both HPC and AI workloads—bringing together CPU, GPU, memory, and interconnect into a tightly integrated system.

Figure 1. The Lenovo ThinkSystem SC777 V4 Neptune Server (AI Training configuration)

The SC777 V4 leverages the NVIDIA GB200 NVL4 architecture, combining Grace CPUs and Blackwell GPUs into a single, tightly integrated node design. Unlike traditional architectures that rely on discrete, loosely coupled components, SC777 V4 is engineered as a balanced, system-level platform.

The design of the SC777 V4 eliminates common bottlenecks seen in conventional systems:

• CPUs starved for memory bandwidth
• GPUs waiting on data movement
• Interconnect limiting multi-GPU scaling

As a result, the platform delivers:

• More consistent performance
• Higher system efficiency
• Better scaling across HPC and AI workloads

Each SC777 V4 server tray includes:

• Compute
  • 2x NVIDIA Grace CPUs (72 Arm Neoverse V2 cores)
  • 4x NVIDIA B200 GPUs (GB200-class modules)
• Memory
  • 960GB of LPDDR5x ECC Memory (480GB per CPU)
  • 744GB HBM3e Memory (186GB per GPU)
• High-speed Interconnect
  • 300 GB/s bi-directional NVLink-C2C connectivity between CPUs
  • 450 GB/s bi-directional NVLink-C2C connectivity between each CPU and its two integrated GPUs, unified coherent memory access
  • 600 GB/s bi-directional NVLink connectivity between each of the four GPUs
• Storage & I/O
  • Supports up to 10x E3.S SSDs (internal)
  • Support for NVIDIA NDR/XDR networking via CPU PCIe slots or GPU-direct modules, up to 800 Gbps per GPU with GPUDirect

Features of the SC777 V4 that drive performance at scale include:

• High-bandwidth memory architecture feeds both CPU and GPU workloads efficiently
• CPU-to-GPU unified, coherent memory enables:
  • Faster data access
  • Reduced need for explicit data movement
  • Improved GPU utilization
• High-speed GPU-to-GPU interconnect:
  • Enables GPUs to operate as a single large compute domain
  • Critical for AI training and distributed HPC workloads
• Integrated into the ThinkSystem N1380 Neptune enclosure (13U)
• Designed for dense scaling:
  • Up to 96 GPUs per standard 19” width rack
• Flexible networking:
  • Supports high-speed NVIDIA NDR and XDR fabrics

The SC777 V4 compute nodes are installed vertically in the ThinkSystem N1380 enclosure, as shown in the following figure.

Figure 2. Front view of the N1380 enclosure with 8x SC777 V4 servers

The SC777 V4 incorporates Lenovo’s Neptune liquid cooling technology, featuring a 100% direct water-cooled design that eliminates traditional airflow constraints and supports high power-density deployments.

• Enables higher sustained system performance under load
• Improves energy efficiency in dense GPU environments
• Supports inlet water temperatures up to 45°C, allowing some data centers to operate without chilled water systems, reducing overall energy consumption

SC777 V4 is not just a GPU-accelerated system - it is a purpose-built platform designed to unify HPC simulation and AI workloads on a single architecture.

With its tightly integrated NVIDIA Grace + Blackwell architecture, the SC777 V4 is built to support a wide spectrum of modern workloads—spanning traditional HPC, enterprise AI, and emerging AI-driven scientific workflows.

• Traditional HPC simulation and modeling
  • Computational fluid dynamics (CFD)
  • Molecular dynamics
  • Electronic structure and quantum chemistry
  • Weather and climate modeling
  • Finite element analysis (FEA)
• Artificial intelligence and data-driven workloads
  • Large language model (LLM) training and fine-tuning
  • AI inference at scale
  • Computer vision and generative AI
  • Graph analytics
• AI for Science and hybrid HPC + AI workflows
  • AI-assisted simulation and surrogate modeling
  • Drug discovery and life sciences research
  • Materials science and quantum modeling
  • Climate prediction with machine learning augmentation
  • Real-time simulation and AI inference pipelines

Functional performance tests are among the first validation steps performed after system installation and are used to baseline compute, memory, and interconnect behavior. These tests are not intended to represent real-world workloads. Instead, they provide a controlled way to verify that each component of the system is delivering expected performance and that the platform is properly optimized. These results also help inform expected behavior for real-world workloads when combined with application profiling data.

To establish these baselines, the following set of well-known synthetic benchmarks are typically used:

• Compute Foundational:
  • HPL (High Performance Linpack)
  • HPL-MxP (mixed precision)
  • GEMM / cuBLAS-based tests
• Memory Bandwidth Foundational:
  • STREAM
  • HPCG (compute + memory behavior)
• Interconnection Foundational:
  • NCCL (GPU-to-GPU communication)
  • NVBandwidth (CPU-to-GPU bandwidth)

The table below lists the functional performance baseline for the SC777 V4 platform.

The synthetic benchmark results are reported using three reference points:

• Results represent the measured performance achieved on the tested SC777 V4 configuration
• Peak represents the theoretical maximum performance obtainable from the underlying hardware based on architectural specifications
• Expected reflects the practical performance range anticipated for a well-configured and properly optimized system.

Comparing measured results against expected and peak values helps validate system health, identify potential performance bottlenecks, and confirm that the platform is operating as intended.

Expected values are derived from publicly available data and internal performance projections informed by prior-generation results and published technical specifications.

Overall, the results show that the system is operating at or near (within 5%) expected performance levels across all major subsystems.

While synthetic benchmarks are essential for validating system readiness, the true measure of a platform is how it performs on real-world workloads.

The results listed in the table below provide an early view of application-level performance across a representative set of HPC and AI workloads. Unlike synthetic benchmarks, these workloads exercise complex compute patterns, memory access behavior, and communication requirements that more closely reflect production environments. As such, they provide valuable insight into how the SC777 V4 performs across a diverse range of scientific computing and AI applications while highlighting the benefits of the NVIDIA GB200 NVL4 architecture.

The table presents an early view of application-level performance on the SC777 V4. Uplift represents the relative performance improvement compared to the baseline system, where 1.50x indicates a 50% improvement and 2.00x indicates twice the performance of the baseline.

Because HPC applications are often bound by memory access, data movement, communication patterns, and FP64 precision requirements, they generally realize more modest generational performance gains than AI workloads.

The SC777 V4 demonstrates strong, balanced performance across a diverse set of real-world workloads. HPC applications see steady, meaningful improvements driven by system balance and memory efficiency. AI workloads realize significantly larger gains, reflecting the rapid evolution of GPU architectures toward data-driven and mixed-precision computing.

The SC777 V4 delivers strong and consistent performance across both HPC and AI workloads. It provides meaningful gains for traditional HPC applications, while AI training and inference workloads benefit from significantly higher improvements. This reflects a broader industry trend, where modern GPU architectures are increasingly optimized for AI and mixed-precision compute. At the same time, the SC777 V4 maintains the performance, stability, and scalability required for demanding scientific workloads.

The SC777 V4 provides a single platform capable of supporting today’s HPC workloads and enabling the next wave of AI-driven innovation.

The benchmark results presented in the document reflect a snapshot of performance under the tested configuration. Results may vary based on system settings, including operating system, BIOS/UEFI configuration, firmware levels, software environment, and workload conditions.

Kevin Dean is the Senior Manager of the HPC/AI Solution Architect Team in the ISG Offerings Group at Lenovo. Kevin oversees the strategy and processes for HPC and AI solutions in this position, contributes his knowledge of HPC/AI application performance, and serves as the CAE/manufacturing architect. Kevin has over 20 years’ experience in HPC, AI, and computational engineering, including nine years at Lenovo focused on HPC/AI performance and solution architecture, and 12 years in aerodynamic design and CFD for the U.S. defense and automotive racing sectors. He earned his Master’s Degree in Aerospace Engineering at the University of Florida, following a Bachelor’s in the same field from Virginia Polytechnic Institute and State University.

Conor Elrick is an HPC/AI Performance Engineer in the ISG Offerings Group at Lenovo working as part of the HPC/AI Solutions Architect Team. Conor performs setup and performance benchmarking on compute servers, networking infrastructure and storage solutions to push hardware to its limits and get the best results for real world scientific workloads. He has been at Lenovo for over 2 years. Conor earned his PhD in Theoretical Particle Physics from the University of Edinburgh in 2024 and a Master’s Degree and Bachelor’s Degree in the same field before that, with a focus on physics simulations on compute systems.