Principal SoC Performance Architect-Microbenchmarks

Key Responsibilities

• Analyze and optimize performance of DCGPU systems across AI training, inference, and HPC workloads
• Identify bottlenecks across hardware, firmware, drivers, runtime, libraries, and applications
• Perform deep kernel-level and system-level profiling to understand performance behavior
• Provide actionable insights to architecture, software, and design teams to improve performance
• Design and develop targeted microbenchmarks to characterize GPU subsystems (compute, memory, interconnect, collectives)
• Build representative system-level workloads reflecting real-world AI/HPC use cases
• Ensure microbenchmarks correlate to application-level performance and architectural intent
• Enable performance validation in pre-silicon environments (simulation/emulation/models)
• Correlate performance data across pre-silicon models and post-silicon measurements
• Develop methodologies to reuse workloads and microbenchmarks across the full lifecycle
• Support bring-up and early silicon performance characterization
• Work across the entire software stack: compiler, runtime, libraries, drivers, and firmware
• Collaborate with ROCm / AI frameworks / kernel teams to improve performance
• Analyze interactions between workload characteristics and hardware execution
• Optimize key kernels (e.g., GEMMs, collectives, attention) and system-level behavior
• Develop and enhance performance measurement, profiling, and analysis tools
• Build automation for performance regression tracking and reporting
• Contribute to unified infrastructure spanning pre-silicon and post-silicon environments
• Cross-functional collaboration with SoC architecture, GPU IP, software, and system teams
• Influence design decisions using data-driven performance insights
• Collaborate with competitive analysis teams to understand gaps vs. industry platforms
• Develop strong intuition and/or models for performance scaling and limits
• Translate performance data into architectural feedback for future GPU designs
• Support competitive benchmarking and performance projections

Preferred Experience

• 10–15+ years of experience in performance engineering for GPUs, HPC systems, or highly parallel SoCs
• Strong understanding of GPU architecture, parallel computing, and memory hierarchies
• Experience with microbenchmark development and system-level workload analysis
• Hands-on experience with performance profiling tools (rocprof, Nsight, perf, etc.)
• Experience analyzing AI/HPC workloads (LLMs, training, inference, communication libraries like RCCL/NCCL)
• Strong background in hardware/software co-design and performance optimization
• Familiarity with pre-silicon (simulation/emulation/models) and post-silicon performance workflows
• Programming expertise in C/C++, Python; experience with GPU programming models (HIP, CUDA, OpenCL)
• Strong analytical and debugging skills with a data-driven mindset
• Experience working across full software stack (compiler → runtime → kernels → system)
• Exposure to performance modeling, scaling analysis, or competitive benchmarking is a plus