SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Graphcore, part of SoftBank Group, is a leading innovator in AI computing hardware, software, and data center infrastructure. As Principal Storage Architect, you will define the high-performance storage architecture for Graphcore's AI computing and data center infrastructures, leading the design of local and distributed storage tiers that provide the throughput, availability, and predictable tail latency required for large-scale training and inference.
Working within Advanced Architecture, you will address storage requirements for saving model states, offloading key-value caches, large datasets, and high-speed data-loading pipelines. You will combine hands-on performance engineering with Principal-level technical leadership across hardware, software, networking, systems engineering, automation, supply chain, and external technology partners.
Key responsibilities include:
- Define end-to-end architecture and technology roadmap for local, disaggregated, and distributed storage across Graphcore AI server and data center platforms
- Design storage topologies optimizing PCIe lane allocation, network-domain placement, and data paths among CPUs, AI accelerators, memory, and storage systems
- Lead architecture of storage control-plane and data-plane solutions, evaluating commercial and open technologies
- Profile and tune the Linux storage stack, block layer, I/O schedulers, direct I/O paths, file systems, and drivers to improve IOPS, bandwidth, and 99.99th-percentile latency
- Optimize storage for AI workloads including model checkpointing, key-value cache offload, data ingestion, and direct data movement between NVMe storage and accelerator memory
- Set technical direction for NVMe SSD lifecycle management across E1.S, E3.S, U.2, and U.3 devices
- Define telemetry and alerting requirements for direct-attached and distributed storage
- Provide architecture requirements and technical guidance to automation teams
- Lead root-cause analysis for complex storage failures and performance degradation
- Partner with storage vendors, supply chain, networking, systems engineering, and product teams
- Evaluate emerging storage, interconnect, and memory-tiering technologies
Requirements:
- Bachelor's degree (or equivalent experience) or master's degree in computer science, computer engineering, information technology, electrical engineering, or related field
- Extensive hands-on experience in storage engineering and architecture for AI, high-performance computing, hyperscale, or large data center environments, with technical leadership at Principal, Staff, or Lead Architect scope
- Deep knowledge of NVMe, PCIe Gen5 or Gen6 architecture, NAND flash behavior, SSD form factors, endurance, firmware, and device lifecycle management
- Deep operational knowledge of Linux storage internals including block layer, I/O schedulers, direct I/O, kernel and driver behavior, and file systems such as XFS, ext4, or ZFS
- Hands-on experience with storage benchmarking and profiling tools such as fio, blktrace, iostat, and perf
- Programming and automation skills in Python and Bash, with experience using structured data formats and REST APIs
- Demonstrated ability to diagnose complex hardware and software interactions including kernel failures, PCIe errors, network issues, performance regressions, and SSD firmware defects
- Experience defining telemetry signals, alert thresholds, dashboards, and operational response criteria for storage health and performance at fleet scale
- Proven ability to lead cross-functional architecture decisions, influence senior engineers and leaders, and communicate technical risks and tradeoffs clearly
Preferred qualifications include experience with GPU Direct Storage or comparable direct accelerator-to-storage technologies, NVMe over Fabrics deployments using RoCEv2 or TCP, Storage Performance Development Kit or other user-space storage frameworks, high-performance parallel and distributed file systems such as Weka, VAST Data, Lustre, DAOS, or DDN, and knowledge of Compute Express Link.
