GPU Architecture and AI Performance Simulation Engineer

Overview

Analyze and model AI workload performance on GPU and SoC architectures to shape architectural decisions for Intel's future AI GPU accelerator roadmap.

What you'll do

• Analyze AI workload behavior on GPU architectures, distinguishing compute-bound, memory-bound, and cache-bound behavior to identify performance bottlenecks and optimization opportunities.
• Build and maintain roofline analytical models and cycle-accurate performance models to support hardware what-if studies and guide micro-architectural decisions.
• Analyze AI workloads built on Transformer and Mixture-of-Experts (MoE) architectures, inference and training, and assess the hardware-utilization impact of kernel-level optimizations such as Flash Attention, Flash Decode, Sparse Attention, and Causal Masking.
• Model and analyze multi-GPU and at-scale AI systems, including workload distribution, synchronization overhead, and communication-compute overlap.
• Drive workload and architectural bottleneck analysis for the early architecture definition phases of Intel's future AI GPU accelerator roadmap products.
• Guide and mentor the team in delivering analytical models and end-to-end system-level performance projections; review modeling methodology and drive consistency across studies.
• Propose new architectural enhancement ideas, backed by quantitative modeling data and workload evidence.
• Maintain a forward-looking view of the AI GPU and accelerator market, tracking competitive roadmaps, whitepapers, and conferences, and translate those trends into architectural requirements while keeping the team updated on applicable AI innovations.

What you'll need

• Bachelor's degree with 10+ years, master's degree with 8+ years, or Ph.D. with 5+ years of experience in System/GPU/SoC performance engineering, AI workload analysis, or hardware architecture.
• Deep knowledge of GPU microarchitecture (NVIDIA SM/CUDA cores, Intel Xe cores, or equivalent), GPU execution pipelines, and cache hierarchy design (L1/L2/L3).
• Familiarity with SoC design principles, on-chip fabric, memory subsystems, and interconnects.
• Experience with multi-GPU system architectures (NVLink, PCIe, CXL, or equivalent) and scaling AI workloads across multiple devices.
• Good understanding of distributed training/inference techniques: Tensor Parallel, Data Parallel, Pipeline Parallel, Expert Parallel, and Model Split methodologies.

Nice to have

• Experience with cycle-accurate simulators or hardware performance emulation platforms.
• Familiarity with AI framework internals (PyTorch, JAX, TensorFlow) and kernel-level optimization (CUDA, SYCL, HIP, Triton).
• Exposure to LLM inference optimization techniques (quantization, KV-cache management, speculative decoding).
• Experience with performance profiling tools (e.g., NVIDIA Nsight, Intel VTune, AMD ROCProfiler, or similar).
• Publications or patents in GPU architecture, AI systems performance, or related areas.

Details

• Location: India, Bangalore.
• Work mode: hybrid, splitting time between on-site at the assigned Intel site and off-site.
• Shift: Shift 1 (India).
• Job type: Experienced Hire.