Skip to main content

PROJECT / JAPAN / OSAKA / GENOMICS

Human genome research infrastructure.

A large-scale AI and high-performance computing platform designed for genomic sequencing, analysis and research collaboration.

The project combined an 800 Gb/s InfiniBand compute fabric, high-speed Ethernet storage networking and Pure Storage data platforms to create a balanced research infrastructure for data-intensive genomics workloads.

Explore Architecture Back to Projects

PROJECT OVERVIEW

Infrastructure designed for data-intensive genomic research.

The architecture was structured around large parallel datasets, GPU-accelerated analysis, high-throughput storage access and low-latency communication between compute nodes.

~1,000800 Gb/s InfiniBand compute-facing ports
~1,000NVIDIA B300 GPUs in the conceptual full-scale configuration
125Eight-GPU B300 compute systems
21Quantum-X800 switches in a two-level fabric
800GEnd-to-end InfiniBand connectivity per compute port
Architecture figures are presented as an approximate project-scale configuration for portfolio communication. Final counts depend on the exact server, rail, redundancy, storage and management design.

800G INFINIBAND FABRIC

Non-blocking two-level fat-tree topology.

The conceptual fabric uses NVIDIA Quantum-X800 switches with 144 ports of 800 Gb/s connectivity.

Fourteen leaf switches provide 72 compute-facing and 72 spine-facing ports each. Seven spine switches aggregate 1,008 leaf uplinks, producing a balanced non-blocking topology for approximately 1,000 compute ports.

14 leaf switches 7 spine switches 144 × 800G ports per switch 1,008 compute-facing ports 1,008 spine-facing links Adaptive routing and SHARP

COMPUTE PLATFORM

B300 systems aligned directly with the network fabric.

Each eight-GPU B300 system contributes eight ConnectX-8 800G ports, allowing a direct conceptual mapping of one high-speed fabric port per GPU.

GPU Compute Layer

Approximately 125 eight-GPU B300 systems provide around 1,000 Blackwell Ultra GPUs for genomics pipelines, machine learning, variant analysis and large-scale data processing.

  • 8 B300 GPUs per system
  • 2.1 TB GPU memory per system
  • 8 × ConnectX-8 ports
  • 800G per compute port
  • NVLink inside each node
  • InfiniBand between nodes

Research Workloads

The platform supports a mix of genomics and AI workflows rather than one isolated benchmark.

  • Sequence alignment
  • Variant calling
  • Population-scale analysis
  • Deep learning models
  • Clinical research pipelines
  • Data-intensive collaboration

PURE STORAGE / DATA FABRIC

Fast storage communication without slowing the compute fabric.

The storage architecture was separated from the core InfiniBand compute fabric and optimized around a dedicated high-speed Ethernet data network.

Pure Storage FlashBlade systems provided scale-out file and object access for genomic datasets, research pipelines, shared analytics and model data.

Gateway and service nodes connected the InfiniBand compute environment with the Ethernet storage fabric, allowing the architecture to use each network for the workload it handles best.

Pure Storage FlashBlade High-speed Ethernet fabric InfiniBand compute backbone Gateway and service nodes Shared genomic datasets File and object access Checkpoint and pipeline data Independent scaling

PROJECT OUTCOMES

A unified platform for genomics, AI and research data.

The project connected compute, networking and storage as one research platform rather than as isolated infrastructure components.

Parallel genomics analysisLarge numbers of samples can be processed across many GPU nodes.
Low-latency GPU communication800G InfiniBand supports distributed training and tightly coupled computation.
High-throughput data accessPure Storage and dedicated Ethernet prevent storage traffic from competing with the compute fabric.
Scalable research platformCompute, storage and network layers can expand independently as datasets and workloads grow.
Customer name, exact location, final bill of materials and operational details are intentionally generalized for confidentiality. Product names identify the technology class used in the project narrative and do not imply vendor endorsement.

RESEARCH INFRASTRUCTURE

Planning an AI or genomics platform?

We can help structure compute, InfiniBand, high-speed Ethernet and storage as one balanced architecture.

Discuss Your Project View All Projects