AI Storage Architect: NVMe/PCIe Expert for AI Data Centers

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 and proper deployment of storage components into our AI server fleet. “Futures”: As a storage architect you will be responsible for scoping the industry for new storage technologies that will give us an advantage on future AI Server designs. Storage Control and Dataplane Definition: Storage expert in charge of the definition and architecture of our external storage Data plane and Control plane solutions using leading vendor solutions such as Weka, VAST, Pure Storage, and others. REQUIREMENTS Bachelor’s, master’s degree or equivalent experience in Information Technology, Computer Science, Computer engineering, or a related field. 12+ years of hands-on experience in storage engineering and architecture and 5+ years supporting storage solutions on AI or Hyperscale Datacenters. Deep NVMe/PCIe Knowledge: Expert-level understanding of the NVMe protocol, PCIe architecture (Gen 5/Gen 6), NAND flash characteristics, and SSD controller behavior (garbage collection, wear leveling, write amplification). Linux Storage Stack: Deep operational knowledge of Linux internals, specifically the block layer, I/O schedulers, direct I/O, and modern local file systems (XFS, ext4, ZFS). Performance Profiling: Extensive hands-on experience with benchmarking and profiling tools such as fio, blktrace, iostat, and perf to simulate AI workloads and isolate bottlenecks. Scripting: Strong programming skills in Python, JSON, Bash or related scripting languages for building validation scripts, automation pipelines, and interacting with REST APIs. Root Cause Analysis: Proven track record of debugging complex hardware/software interactions, including kernel panics, PCIe bus errors, and SSD firmware bugs. Telemetry: Point of contact responsible for defining the telemetry sensors and rules to activate alerts in our Datacenter fleet monitoring frameworks. DIFFERENTIATORS GPU Direct Storage (GDS): Experience implementing, testing, and troubleshooting GDS or similar technology for direct GPU-to-storage communication. NVMe over Fabrics (NVMe-oF): Experience designing or managing NVMe-oF (RoCEv2 or TCP) deployments to disaggregate storage without sacrificing local-NVMe performance. Storage Software Development: Experience with the Storage Performance Development Kit (SPDK) or developing user-space storage applications. Distributed AI File Systems: Familiarity with high-performance parallel file systems commonly used in AI/HPC clusters (e.g., Weka, VAST Data, Lustre, DAOS, or DDN). Emerging Tech: Knowledge of Compute Express Link (CXL) and its implications for future memory and storage tiering in AI servers. We welcome people of different backgrounds and experiences and are committed to building an inclusive work environment that makes Graphcore a great home for everyone. We are an equal opportunity employer and want to build a work environment where everyone is happy, productive and respectful so they can do their best work. If you have a disability or additional need that requires accommodation, just let us know. #J-18808-Ljbffr Cerebras