About Modular
At Modular, a Qualcomm company, we’re on a mission to revolutionize AI infrastructure by systematically rebuilding the AI software stack from the ground up. Our team, made up of industry leaders and experts, is building cutting-edge, modular infrastructure that simplifies AI development and deployment. By rethinking the complexities of AI systems, we’re empowering everyone to unlock AI’s full potential and tackle some of the world’s most pressing challenges.
If you’re passionate about shaping the future of AI and creating tools that make a real difference in people’s lives, we want you on our team. You can read about our culture and careers to understand how we work and what we value.
About the role:
We are looking for an AI Performance Engineer to join the ASIC Performance team. You will design, implement, and tune high-performance attention kernels—and a broader set of foundational AI kernels—for new accelerator architectures.
Attention is evolving quickly across models and hardware: grouped-query and multi-head attention, prefill and decode, paged KV caches, sliding-window and global attention, and new sparse or fused variants each create different compute, memory, and data-movement tradeoffs. In this role, you will translate those algorithmic requirements into production-quality implementations that exploit the unique capabilities of GPUs, NPUs, ASICs, and other emerging accelerators.
You will also contribute to the architecture and development of Modular’s Attention framework, building reusable abstractions that let model and kernel developers express attention variants cleanly while supporting hardware-specific specialization and excellent end-to-end performance.
This is a senior individual-contributor role for an engineer who can independently own ambiguous performance problems, collaborate across organizational boundaries, and turn research ideas and hardware capabilities into robust product technology. Our goal is to deliver performant tokens to end cloud users with production-ready workloads, not just microbenchmarks.
LOCATION: Candidates based in the US, Canada, and UK are welcome to apply. You can work in our office in Los Altos, CA, or Edinburgh, or remotely from home. Onboarding for new hires is conducted in-person at the appropriate office.
What you will do:
- Design and implement high-performance attention kernels for prefill and decode, including variants such as MHA, GQA, paged attention, sliding-window/global attention, and fused FlashAttention-style algorithms.
- Build and optimize general AI kernels—including matrix multiplication, softmax, normalization, activation, embedding, reduction, and data-movement primitives—needed to bring up modern models on new accelerators.
- Contribute to the design and implementation of the Attention framework, with reusable interfaces that balance portability, composability, tunability, and hardware-specific optimization.
- Profile workloads, identify bottlenecks, establish performance models, and tune implementations against hardware limits and competitive baselines.
- Develop benchmarks, correctness tests, regression tests, and automated performance analysis to ensure kernels remain reliable and fast across model shapes and hardware generations.
- Partner with the rest of Modular teams to contribute to overall software stack improvements and share learnings.
- Work directly with hardware partners when needed to understand new architectures, validate toolchains, diagnose low-level issues, and influence hardware/software interfaces.
- Write design documents, participate in technical reviews, and share architecture and performance insights across the team.
- Mentor teammates and raise the engineering bar for kernel quality, performance methodology, and maintainable low-level software.
What you bring to the table:
- 5+ years of relevant industry or research experience in high-performance computing, AI kernel development, accelerator programming, compiler engineering, or a closely related field.
- Demonstrated experience writing and optimizing production-quality GPU or accelerator kernels, with a strong understanding of parallel algorithms, numerical behavior, memory access patterns, synchronization, and data movement.
- Strong knowledge of modern attention algorithms and their implementation tradeoffs, including the distinct performance characteristics of prefill and decode.
- Strong understanding of architectures beyond conventional GPUs—such as NPUs, DSPs, TPUs, or custom ASICs—and the ability to reason about unfamiliar compute units, memory hierarchies, and execution models.
- Experience with at least one heterogeneous programming model or kernel ecosystem such as CUDA, Triton, SYCL, OpenCL, HIP, or a vendor accelerator SDK.
- Strong performance-analysis skills and proficiency with relevant profiling, tracing, benchmarking, and debugging tools.
- Ability to read architecture manuals, compiler output, and low-level generated code; form hypotheses from data; and systematically close performance gaps.
- A collaborative, team-oriented approach; clear written and verbal communication; intellectual curiosity; and comfort owning ambiguous technical problems from investigation through productization.
Helpful but not required
- Knowledge of the Mojo programming language or experience authoring kernels and performance-sensitive libraries in Mojo.
- Knowledge of MLIR, LLVM, and related AI compiler technologies, including dialect design, lowering pipelines, code generation, scheduling, or autotuning.
- Experience implementing FlashAttention or other fused attention algorithms, custom attention operators, paged KV-cache kernels, or inference-specific attention optimizations.
- Experience with modern kernel DSLs and libraries such as CuTe/CUTLASS, Triton, Pallas, or similar systems.
- Familiarity with AI framework internals and operator integration in systems such as PyTorch, JAX, TensorFlow, vLLM, SGLang, or TensorRT-LLM.
- Experience bringing up a model or kernel library on a new hardware platform, including working through compiler, runtime, driver, and hardware constraints.
- Experience with performance modeling, autotuning, numerical validation, or benchmarking infrastructure.
- Familiarity with LLM architectures such as GPT and Gemma, including grouped-query attention, mixture-of-experts, speculative decoding, and long-context inference.
- An advanced degree in Computer Science, Electrical Engineering, Computer Engineering, or a related field.
What Modular brings to the table:
- Amazing Team. We are a progressive and agile team with some of the industry’s best engineering and product leaders.
- World-class Benefits. In order to attract the best, we need to offer the best. Your benefits package may include comprehensive healthcare coverage, retirement and savings programs, employee stock purchase opportunities, paid time off, wellbeing resources, family support programs, and learning and development opportunities. Please note that specific benefit packages may vary based on your location, you can read more about benefits offered by Qualcomm here.
- Competitive Compensation. We offer very strong compensation packages, including RSU grants. We want people to be focused on their best work and believe in tailoring compensation plans to meet the needs of our workforce.
- Team Building Events. We organize regular team onsites and local meetups in Los Altos, CA as well as different cities. Traveling 2-4 times a year is expected for all roles.
Working at Modular will enable you to grow quickly as you work alongside incredibly motivated and talented people who have high standards, possess a growth mindset, and a purpose to truly change the world.
The estimated base salary range for this role to be performed in the US, regardless of the state, is $180,000 - $270,000 USD.
The estimated base salary range for this role to be performed in Canada, regardless of the province, is $172,400.00 - $258,600 CAD.
The estimated base salary range for this role to be performed in Edinburgh is £99,600–£150,000 GBR.
The salary for the successful applicant will depend on a variety of permissible, non-discriminatory job-related factors, which include but are not limited to education, training, work experience, business needs, or market demands. This range may be modified in the future. The total compensation for a candidate will also include annual target bonus, equity, and benefits, with equity making up a significant portion of your total compensation.
For candidates who fall outside of the listed requirements, we nevertheless encourage you to apply as we may have upcoming openings that are lower/higher level than the ones advertised.
5+ years in high-performance computing or AI kernel development; experience writing production GPU/accelerator kernels; strong knowledge of attention algorithms and non-GPU architectures; proficiency in CUDA, Triton, SYCL, OpenCL, HIP, or vendor SDK; strong performance analysis skills.
Key Responsibilities
- designing kernels
- optimizing performance
- mentoring teammates
Skills & Tools
CUDA, Triton, SYCL, OpenCL, HIP, Mojo, MLIR, LLVM, CuTe, CUTLASS, Pallas, PyTorch, JAX, TensorFlow, vLLM, SGLang, TensorRT-LLM
Job Details
- Category: Software Development
- Seniority: Senior Level
- Commitment: Full Time
- Workplace: Remote — United States or Canada or United Kingdom
- Salary: USD 180,000 – 270,000 / year
- Languages: English
Benefits
- Generous paid time off
- Retirement plan
About Modular
A Qualcomm company building modular AI infrastructure and software to simplify AI development and deployment. — Industry: Information Technology