Graphcore is a globally recognized leader in Artificial Intelligence computing systems. The company designs advanced semiconductors and data centre hardware that provide the specialized processing power needed to drive AI innovation, while delivering the efficiency required to support its broader adoption.
As part of the SoftBank Group, Graphcore is a member of an elite family of companies responsible for some of the world’s most transformative technologies. We are opening a new AI Engineering Campus in Austin, which will play a central role in Graphcore's work building the future of AI computing.
About the Role
We are seeking a Storage Architect to design, evaluate, and optimize the high-performance storage architectures powering our AI data centers. In AI training and inference, keeping GPUs fed with data is one of the most critical challenges; even a slight I/O bottleneck can lead to massive inefficiencies. In this role, you will be our resident expert on solid-state storage, focusing heavily on NVMe SSDs, PCIe topologies, and the Linux storage stack. You will ensure our local and distributed storage tiers deliver the microsecond predictable latency and massive throughput required for large language model (LLM) checkpointing, massive datasets, and high-speed data loading pipelines.
RESPONSIBILITIES
- Hardware Qualification: Lead the evaluation, qualification, and lifecycle management of NVMe SSDs (e.g., E1.S, E3.S, U.2/U.3 form factors) from major vendors for use in high-density AI servers and AI Clusters.
- Architecture & Topology: Design and optimize local server storage topologies, managing PCIe lane distribution, minimizing NUMA node crossings, and ensuring optimal data paths between NVMe drives, CPUs, and GPUs.
- Performance Tuning: Conduct rigorous performance profiling and tuning of the Linux kernel storage stack, block layer, and file systems to maximize IOPS and bandwidth while minimizing tail latency (i.e., 99.99%).
- AI Workload Optimization: Optimize storage configurations specifically for AI workloads, including tuning for GPU direct storage to enable direct memory access between NVMe storage and GPU memory, bypassing the CPU.
- Telemetry & Automation: Drive telemetry strategy for the storage subsystem including SSD health monitoring (i.e., wear leveling, DWPD, thermals) and latency anomaly detection. Provide requirements and technical guidance to the automation team building test framework to characterize the storage subsystem in our AI Platforms. Additionally, manage firmware rollouts at scale.
- Troubleshooting: Act as the highest point of escalation for complex, storage-related performance degradation or hardware failures, utilizing tools like nvme-cli, fio, and PCIe analyzers.
- Vendor engagement: Responsible for working closely with storage vendors, supply chain teams, and our Systems engineering team in the selection a