AI Performance Engineer

This is a fully remote position, open to applicants in United Kingdom.

📋 Description

• Design and develop high-performance attention kernels for prefill and decoding, including MHA, GQA, paged attention, sliding-window/global attention, and fused FlashAttention-style algorithms.

• Create and enhance general AI kernels, covering matrix multiplication, softmax, normalization, activation, embedding, reduction, and data-movement primitives.

• Contribute to the architecture and execution of Modular’s Attention framework.

• Analyze workloads, pinpoint bottlenecks, establish performance models, and optimize implementations based on hardware constraints and competitive benchmarks.

• Develop benchmarks, correctness assessments, regression tests, and automated performance evaluations.

• Collaborate with Modular teams to enhance the software stack and share insights.

• Work alongside hardware partners to understand architectures, validate toolchains, diagnose low-level issues, and influence hardware/software interactions.

• Compose design documentation and engage in technical reviews.

• Share insights on architecture and performance throughout the team.

• Mentor colleagues and enhance kernel quality, performance methodologies, and maintainable low-level software.

• Take ownership of complex performance challenges, transforming research ideas and hardware capabilities into reliable product technology.

⛳️ Requirements

• 5+ years of relevant experience in high-performance computing, AI kernel development, accelerator programming, compiler engineering, or a related discipline.

• Proven experience in writing and optimizing production-quality GPU or accelerator kernels.

• Deep understanding of parallel algorithms, numerical behaviors, memory access patterns, synchronization, and data movement.

• Solid knowledge of modern attention algorithms and their implementation trade-offs, particularly in prefill and decode performance characteristics.

• Strong grasp of architectures beyond standard GPUs, including NPUs, DSPs, TPUs, or custom ASICs.

• Familiarity with at least one heterogeneous programming model or kernel ecosystem like CUDA, Triton, SYCL, OpenCL, HIP, or a vendor-specific accelerator SDK.

• Strong skills in performance analysis with proficiency in profiling, tracing, benchmarking, and debugging tools.

• Ability to interpret architecture manuals, compiler outputs, and low-level generated code.

• Capacity to develop hypotheses from data and systematically address performance discrepancies.

• A collaborative mindset, team-oriented approach, excellent written and verbal communication skills, intellectual curiosity, and comfort in tackling ambiguous technical issues.

• Familiarity with the Mojo programming language is advantageous but not essential.

• Understanding of MLIR, LLVM, and associated AI compiler technologies is beneficial but not mandatory.

• Experience in implementing FlashAttention or other fused attention algorithms is a plus but not required.

• Familiarity with modern kernel DSLs and libraries such as CuTe/CUTLASS, Triton, Pallas, or similar systems is a bonus but not required.

• Acquaintance with AI framework internals like PyTorch, JAX, TensorFlow, vLLM, SGLang, or TensorRT-LLM is helpful but not mandatory.

• Experience in initiating a model or kernel library on a new hardware platform is advantageous but not essential.

• Knowledge in performance modeling, autotuning, numerical validation, or benchmarking infrastructure is helpful but not required.

• Familiarity with LLM architectures such as GPT and Gemma is a plus but not essential.

• An advanced degree in Computer Science, Electrical Engineering, Computer Engineering, or a related field is beneficial but not essential.

🏝️ Benefits

• Comprehensive healthcare coverage.

• Retirement and savings programs.

• Employee stock purchase opportunities.

• Paid time off.

• Wellbeing resources.

• Family support programs.

• Learning and development opportunities.

• RSU grants.

• Annual target bonus.

• Equity.

• Regular team onsite events and local meetups.

• Travel 2–4 times a year.

• Option for remote work from home.

• In-person onboarding at the designated office.