Solution Architect - GPU/TPU Kernel Optimization

EPAM Systems; Chennai District; On-site; INR 2,000,000 - 3,000,000; Full time

Job summary

EPAM Systems is looking for an experienced Solution Architect in Chennai, India, specializing in GPU/TPU Kernel Optimization. The successful candidate will design and optimize high-performance kernels for machine learning operations, redefining performance boundaries while shaping developer infrastructure. The role requires 12-18 years of software development experience and expertise in optimizing code using languages like Pallas, CUDA, and Triton. Excellent communication skills and knowledge of ML frameworks such as JAX and PyTorch are essential.

Qualifications

• 12-18 years of experience in software development.
• Expertise in optimizing TPU/GPU code using low-level kernel languages.
• Strong knowledge of ML frameworks, model optimization, and low-precision formats.

Responsibilities

• Design and optimize high-performance kernels targeting TPU and GPU architectures.
• Architect infrastructure for benchmarking suites and performance analysis tools.
• Engage with ML researchers and framework developers to enhance adoption.

Skills

GPU/TPU Kernel Optimization

Pallas language

Triton

CUDA

JAX

PyTorch

Performance Optimization

Cross-Functional Communication

Tools

MLIR

OpenXLA

Job description

We are seeking an experienced Solution Architect specializing in GPU/TPU Kernel Optimization to design and optimize high-performance kernels for cutting-edge Machine Learning operations. In this role, you will redefine performance boundaries across massive training runs and high-speed inference workloads while shaping the developer infrastructure that powers next-generation AI systems.

Responsibilities

• Design and optimize high-performance kernels (using languages like Pallas, Mosaic and Triton) targeting Tensor Processing Unit (TPU) and Graphics Processing Unit (GPU) architectures for critical Machine Learning (ML) operations, redefining what’s possible from massive training runs to high-speed inference
• Architect infrastructure such as benchmarking suites, autotuning frameworks, performance analysis tools, regression testing and documentation
• Transform how the developer community interacts with increasingly critical custom kernels in key Open-Source Software (OSS) libraries
• Track the latest advancements in hardware architectures, compiler technologies and AI models to identify new opportunities for performance optimization through custom kernels
• Engage with ML researchers, framework developers (Just After eXecution JAX, PyTorch) and compiler engineers (Accelerated Linear Algebra XLA) to enhance adoption
• Identify new requirements and address bottlenecks by providing appropriate solutions

Requirements

• 12-18 years of experience in software development
• Expertise in optimizing TPU/GPU code using low-level kernel languages like Pallas, Compute Unified Device Architecture (CUDA) or Triton
• Knowledge of ML Frameworks (JAX/PyTorch), common operations like attention and Mixture of Experts (MoEs), including model optimization and low-precision formats
• Understanding of modern accelerators (e.g., data movement, pipelining, heterogeneous compute and scale-out)
• Understanding of compiler principles (optimization, code generation) and toolchains such as MLIR, OpenXLA
• Showcase of building developer infrastructure, including Open-Source Software (OSS) libraries, flexible high-performance APIs and easy-to-consume documentation to empower the community
• Excellent investigative and problem-solving capabilities with communication skills across cross-functional teams