Senior GPU Inference Performance Engineer

Advanced Micro Devices, Inc•Santa Clara, CA

7d•Hybrid

About The Position

We are looking for a Senior GPU Inference Performance Engineer to own end-to-end performance analysis of GPU-accelerated AI inference workloads. You will profile, diagnose, and explain performance across the full stack, from GPU silicon through the software runtime, and drive competitive positioning against other accelerator vendors. This role sits at the intersection of hardware, systems software, and AI serving frameworks, and requires someone who can go deep on a trace and present findings to product and executive stakeholders. A hands-on performance engineer who is equally comfortable reading a GPU trace and briefing executives. You are curious, evidence-driven, rigorous and you don't stop at "X is faster," you explain why, rooted in hardware and software evidence. You collaborate across hardware, systems software, and AI serving framework teams, communicate clearly in written reports and presentations, and thrive at the intersection of silicon, systems, and AI.

Requirements

Background in GPU performance engineering, HPC, or systems performance analysis
Hands-on proficiency with either AMD (ROCm, rocProfiler, Omniperf/Omnitrace) or NVIDIA (CUDA, Nsight Systems/Compute, NCU) profiling toolchains, with deep understanding of GPU architecture: warp/wavefront execution, memory hierarchy (registers → LDS/shared → L2 → HBM), occupancy, and instruction-level parallelism
Experience profiling vLLM, SGLang, or equivalent LLM serving frameworks, including quantization workflows (FP8, MXFP4, INT4, AWQ, GPTQ) and their performance implications
Experience with multi-GPU and multi-node inference — tensor parallelism, pipeline parallelism, or PD disaggregation over RDMA/RoCE — including RCCL/NCCL profiling and network tools (perftest, ib_write_bw, tcpdump, Memory Fabric counters)
Demonstrated ability to explain performance differences in written reports or presentations — not just "X is faster" but why, rooted in hardware and software evidence
Strong Python and C/C++ skills; comfort reading GPU kernel code (HIP/CUDA)
Experience with Kubernetes GPU scheduling, MIG, and GPU operator performance, or contributions to open-source inference or profiling projects

Responsibilities

Full-stack GPU profiling: Instrument and analyze inference workloads across AMD Instinct (ROCm, rocProfiler, Omniperf) and NVIDIA (CUDA, Nsight Systems/Compute, DCGM) GPUs. Identify bottlenecks spanning HBM bandwidth, compute utilization, kernel scheduling, memory allocation, and PCIe/Infinity Fabric data movement.
AI serving framework performance: Profile and optimize inference engines including vLLM, SGLang, and emerging serving runtimes. Understand KV-cache management, continuous batching, PagedAttention, speculative decoding, and quantization (FP8, MXFP4, INT4) effects on throughput and latency.
Competitive performance analysis: Design and execute head-to-head benchmarks (AMD vs. NVIDIA) on standardized LLM workloads. Produce clear, data-backed explanations of why performance differs — attributing gaps to specific hardware features (HBM bandwidth, compute density, interconnect topology), software maturity (kernel libraries, operator fusion, graph compilation), or configuration differences.
Multi-server inference networking: Profile and optimize distributed inference topologies including prefill-decode (PD) disaggregation, pipeline parallelism, and tensor parallelism across multi-node clusters. Analyze network-level bottlenecks using RDMA/RoCE traces, NCCL/RCCL collective profiling, and NIC-level counters (Pensando, ConnectX). Quantify the impact of network latency, bandwidth, and congestion on end-to-end inference SLAs.
GPU operator and Kubernetes stack: Profile the overhead introduced by GPU operators, device plugins, container runtimes (Docker, containerd), and Kubernetes scheduling on inference latency. Identify and resolve jitter, cold-start, and resource contention issues in production serving environments.
Tooling and automation: Build reproducible benchmarking harnesses, profiling scripts, and performance regression dashboards. Automate trace collection and analysis to support continuous performance validation across driver, firmware, and framework updates.