Research Scientist - AI Compute & DPU

We are looking for talented individuals to join our team. As a graduate, you will get opportunities to pursue bold ideas, tackle complex challenges, and unlock limitless growth. Successful candidates must be able to commit to an onboarding date by the end of the year. Please state your availability and graduation date clearly in your resume. Team Introduction: The ByteDance DPU (Data Processing Unit) team builds foundational cloud and AI computing infrastructure for ByteDance and Volcano Engine. Our mission is to advance the architecture, development, and research of next-generation software-hardware co-design technologies across compute, networking, and storage for cloud and AI computing. Our technology stack spans - Cloud virtualization, hypervisors, and operating systems - High-performance networking, including DPDK and RDMA - High-speed interconnects, virtual switching, and network offload - Distributed storage and I/O acceleration - Orchestration and scheduling for AI/ML workloads Our organization—AI Compute + DPU—unifies three foundational layers: - AI Compute & Orchestration Planet-scale, cloud-native scheduling and serverless systems built on Kubernetes, operating across hundreds of clusters with hundreds of millions of containers and jobs daily. - Inference Infrastructure We support Seed business workloads (including SeedDance and other ByteDance models), as well as a wide range of open-source models. Through the open-sourcing of AIBrix, we are re-architecting LLM inference around KV cache systems, multimodal serving, advanced scheduling, and disaggregated execution, pushing the frontier of performance, cost efficiency, and latency. - DPU & Hardware-Accelerated Systems Next-generation software–hardware co-design across compute, networking, and storage—covering GPU virtualization, RDMA/DPDK-based networking, high-speed interconnects, and distributed storage acceleration. Together, we operate a globally distributed, hyperscale AI infrastructure powering products such as Doubao/Dola and CapCut, with a rapidly growing fleet of accelerators and increasingly deep integration with open-source ecosystems. What Makes This Different This is not an incremental infrastructure team. We are redefining how AI systems are built and operated: - Re-architecting LLM inference for heterogeneous, disaggregated environments - Designing next-generation scheduling and caching primitives for AI workloads - Bridging research breakthroughs with production systems at scale - Building infrastructure that can self-optimize via AI (AI for systems) - Driving cost-efficiency at extreme scale (GPU, CPU, network, power) This is a system-defining opportunity to work on problems that do not yet have established solutions. Who We Are Looking For Through our Global Frontier Recruitment Program, we are selectively hiring a very small number of PhD candidates who demonstrate the potential to: define technical directions, lead complex systems end-to-end, and create lasting impact in AI infrastructure. We are not looking for strong candidates. We are looking for candidates whose absence would be felt. Topic Content: With the large-scale adoption of LLMs and AI agents, traditional cloud-native infrastructure can no longer meet the ultra-high performance and elasticity requirements of AI workloads. This topic conducts systematic research across the entire AI infrastructure stack: 1. Network and Observability: Research intelligent fault localization and root cause analysis for large-scale AI clusters, combined with intelligent tuning of time-series databases to improve cluster stability. 2. Storage Systems: Develop serverless high-performance elastic file systems and storage acceleration architectures specifically for AI scenarios, explore hardware-software co-optimization for DPU, and overcome AI storage performance bottlenecks. 3. Data Center Power Scheduling: Research GPU/CPU/MEM heterogeneous collaborative scheduling technologies, build a heterogeneous power orchestration system for AI agents, and address scheduling challenges including heterogenous workloads and state dependencies. 4. Vector Retrieval: Optimize core vector retrieval technologies for LLM-powered applications, building a cloud-native distributed vector index engine to meet ultra-large-scale vector retrieval demands with low latency and low cost. 5. Intelligence and Agent Architecture: Explore automatic infrastructure optimization based on AI Agent workflows, build a self-evolvable business agent framework, and enable full-stack intelligent optimization through AI for Infra. This topic aims to build a next-generation AI-native infrastructure to support the deployment of LLMs and AI agents, improve resource utilization, reduce costs, support elastic scaling, and drive the technological evolution of AI infrastructure. Minimum Qualifications: - Individuals who are completing or recently completed a PhD in Software Development, Computer Science, Computer Engineering, or a related technical discipline, with a focus on distributed and ML systems. - Proven first-author publications in top venues (e.g., OSDI, SOSP, NSDI, NeurIPS, MLSys, etc.) with clear technical contributions. - Strong system-building ability, with hands-on experience implementing or optimizing real systems beyond prototypes. Solid understanding of compute, network architecture, and operating systems. - Deep expertise in at least one of the following: LLM inference / AI-ML systems, system optimization for AI (e.g., scheduling, observability, resource management, high-performance networking), or software-hardware co-design. - Demonstrated ownership of significant technical work (research or systems), with the ability to clearly articulate the problem defined, decisions made, and why the outcome depended on their contribution. - Strong programming skills and the ability to reason about complex system design. Preferred Qualifications: - A sustained track record of high-impact research, with evidence of influence (citations, follow-up work, or adoption in real systems). - Experience taking ideas from research to production, including deployment, evaluation, and iteration in real environments. - Deep expertise in frontier areas such as large-scale model training and inference, heterogeneous computing optimization across GPUs/DPUs/accelerators, software-hardware co-design (e.g., FPGA/ASIC) across networking/storage/distributed compute, or high-performance networking (e.g., RDMA, NCCL, DPDK/SPDK) with hands-on experience in network virtualization (OVS, SR-IOV, eBPF). - Proven ability to lead technical direction, not just contribute—e.g., defining a research agenda, owning a system component, or driving cross-functional decisions. - Contributions to widely used open-source systems or infrastructure projects. - Strong industry research experience or collaboration with leading labs - Ability to reason rigorously about system-level trade-offs (latency, throughput, cost, scalability) and translate them into practical designs.