Systems Performance Principal Engineer, AI Solutions Lab
Read straight from Marvell's careers page, verified 2d ago.Seekless can tailor your résumé to it.
Seekless can write your résumé for this posting first.
Before you apply
SalaryNo pay range posted
LocationSanta Clara, CA
Typefull-time
SenioritySenior
Company size10,000+ people
PostedOct 16, 2026 · 3w ago
Verified live2d ago
At a glanceSummarised by Seekless from the posting.
Experience with AI serving frameworks such as vLLM, SGLang, TensorRT-LLM, or similar platforms
Strong background in RCCL/NCCL, RDMA, communication-compute overlap, and performance profiling
Hands-on experience with any GPU profiling tools (e.g., Nsight Systems/Compute, PyTorch Profiler, or ROCm Compute Profiler) to localize bottlenecks at the kernel level
Demonstrated ability to diagnose end-to-end bottlenecks across application, framework, runtime, and infrastructure layers
Working knowledge of at least one distributed training framework (PyTorch DDP/FSDP, DeepSpeed, or Megatron-LM)
Strong Python and Linux skills
Comfortable working in containerized environments (Docker; Kubernetes a plus)
Ability to translate deep technical analysis into clear recommendations and concise performance narratives
Experience benchmarking distributed large-model inference and training systems
Familiarity with GPU kernel analysis, system tracing, and workload-specific optimization
Familiarity with serving-stack internals such as continuous batching, paged KV-cache attention, and quantization techniques (e.g., GPTQ, SmoothQuant)
Exposure to customer proofs of concept, competitive benchmarking, or externally consumed performance reporting
Go experience is a plus
Skills
vLLM
SGLang
TensorRT-LLM
RCCL
NCCL
RDMA
Nsight Systems
Nsight Compute
PyTorch Profiler
ROCm Compute Profiler
PyTorch DDP
FSDP
DeepSpeed
Megatron-LM
Python
Linux
Go
Docker
Kubernetes
GPTQ
SmoothQuant
About Marvell
Marvell’s semiconductor solutions are the essential building blocks of the data infrastructure that connects our world. Across enterprise, cloud and AI, and carrier architectures, our innovative technology is enabling new possibilities.
At Marvell, you can affect the arc of individual lives, lift the trajectory of entire industries, and fuel the transformative potential of tomorrow. For those looking to make their mark on purposeful and enduring innovation, above and beyond fleeting trends, Marvell is a place to thrive, learn, and lead.
Your Team, Your Impact
The Solutions Architecture team sits within Marvell’s Platform organization, which turns advanced silicon into deployable, customer-ready AI and cloud infrastructure. This team — the AI Solutions Lab — owns the validation layer that proves our platforms perform as promised, across compute, networking, and memory fabrics. Without it, Marvell could design great silicon but could never prove, at rack scale, that it performs the way customers expect — and hyperscalers don’t buy chips on paper, they buy validated systems.
Day to day, the team runs AI benchmarking experiments that mirror real customer workloads, operates GPU/HPC lab clusters, validates Marvell’s Teralynx switching and AI fabrics under production-like conditions, and builds live demos for trade shows — carrying platform programs from concept to production.
What sets this team apart is breadth of ownership: engineers here aren’t limited to validating a single switch chip — they work across the full breadth of Marvell’s AI platform portfolio, from Teralynx network fabrics to integrated optical platforms like co-packaged optics (CPO) and near-packaged optics (NPO), as well as emerging memory fabric products such as CXL and photonic fabric memory appliances. This cross-domain ownership — networking, optics, and memory, together — is work that’s often siloed across separate teams elsewhere, and the results feed directly into customer proof points and executive reviews.
What You Can Expect
• Drive end-to-end performance optimization across AI serving and training stacks using leading models and customer-relevant configurations.
• Identify bottlenecks across the full data path, from application request flow through compute, memory, communication layers, GPU kernels, network transport, and data/storage pipelines.
• Develop and execute benchmark methodologies that are reproducible, technically credible, and relevant to customer use cases.
• Profile serving-stack internals — quantization, scheduling, KV-cache management, and disaggregated prefill/decode — to identify concrete optimization targets.
• Analyze results and develop performance narratives that guide kernel, runtime, communication, and system-level optimizations.
• Work closely with NRE / SRE and network or fabric architects to close competitive gaps through targeted tuning.
• Create performance summaries, visuals, and benchmark artifacts for customer presentations, executive reviews, and trade-show demonstrations.
What We’re Looking For
Required Qualifications
• Experience with AI serving frameworks such as vLLM, SGLang, TensorRT-LLM, or similar platforms.
• Strong background in RCCL/NCCL, RDMA, communication-compute overlap, and performance profiling.
• Hands-on experience with any GPU profiling tools (e.g., Nsight Systems/Compute, PyTorch Profiler, or ROCm Compute Profiler) to localize bottlenecks at the kernel level.
• Demonstrated ability to diagnose end-to-end bottlenecks across application, framework, runtime, and infrastructure layers.
• Working knowledge of at least one distributed training framework (PyTorch DDP/FSDP, DeepSpeed, or Megatron-LM), given the role’s scope spans both training and inference stacks.
• Strong Python and Linux skills; Go experience is a plus.
• Comfortable working in containerized environments (Docker; Kubernetes a plus) to ensure benchmark reproducibility.
• Ability to translate deep technical analysis into clear recommendations and concise performance narratives.
Preferred Qualifications
• Experience benchmarking distributed large-model inference and training systems.
• Familiarity with GPU kernel analysis, system tracing, and workload-specific optimization.
• Familiarity with serving-stack internals such as continuous batching, paged KV-cache attention, and quantization techniques (e.g., GPTQ, SmoothQuant).
• Exposure to customer proofs of concept, competitive benchmarking, or externally consumed performance reporting.