Research Engineer - AI Performance & Kernel Optimization

Zyphra•San Francisco, CA

3d•Onsite

About The Position

Zyphra is an artificial intelligence company based in San Francisco, California. The Role: As a Research Engineer - AI Performance & Kernel Optimization , you will improve and optimize the performance of our large-scale language model training and inference stacks. You will work closely with our pretraining and inference teams to identify bottlenecks, design and implement highly optimized kernels, and push the limits of throughput, latency, and hardware utilization across a range of accelerator platforms. This role is suited for someone who enjoys deep systems work, cares about performance at every level of the stack, and is excited to translate low-level optimizations into meaningful gains for frontier-scale AI systems. You’ll Work Across: Kernel development and optimization for large-scale ML workloads, using any level of the stack from PTX/assembly to CUDA, HIP, Triton, or other GPU DSLs Performance tuning for training and inference stacks across GPUs and other accelerators Profiling and eliminating bottlenecks in memory movement, communication, scheduling, and compute utilization Optimizing distributed training and inference systems for large MoE models, including large-scale model parallelism Portability and optimization across non-NVIDIA hardware, with special interest in AMD hardware such as the MI300x and MI355x Collaboration with research and infrastructure teams to turn systems improvements into real-world model training and inference gains

Requirements

Strong engineering aptitude for building reliable, high-performance systems
Excellent low-level performance intuition and the ability to reason about hardware-software interactions
Are excited to rapidly learn new systems, tools, and hardware environments
Excellent communication and collaboration skills, with the ability to work effectively across research and engineering teams
Enjoy diving deep into the weeds and hunting down the last 10–20% of performance

Nice To Haves

Experience writing highly performant GPU kernels at any level of abstraction–PTX, CUDA, HIP, Triton, or other kernel DSLs
Experience optimizing ML workloads for large-scale training, ideally in language model pretraining or inference environments
Experience with non-NVIDIA accelerator hardware, such as AMD, AWS Trainium, Google TPU, Qualcomm, ARM, Intel, and custom ASICs
Strong understanding of distributed training systems and parallelism schemes, including data parallelism, tensor/model parallelism, pipeline parallelism, sharding, and communication/computation overlap
Experience with performance engineering in other demanding parallel computing environments such as HPC, quantitative finance, scientific computing, graphics, compilers, or numerical simulation
Strong systems intuition around memory hierarchy, bandwidth constraints, kernel fusion, launch overhead, communication overhead, and hardware utilization
Experience using profiling and debugging tools to drive performance improvements
Familiarity with infrastructure underlying large-scale training and inference, including collective communication libraries, and runtime performance analysis
Background in a highly technical field such as physics, mathematics, theoretical computer science, computer science, or electrical engineering
Any HPC experience is a strong plus

Responsibilities

Kernel development and optimization for large-scale ML workloads, using any level of the stack from PTX/assembly to CUDA, HIP, Triton, or other GPU DSLs
Performance tuning for training and inference stacks across GPUs and other accelerators
Profiling and eliminating bottlenecks in memory movement, communication, scheduling, and compute utilization
Optimizing distributed training and inference systems for large MoE models, including large-scale model parallelism
Portability and optimization across non-NVIDIA hardware, with special interest in AMD hardware such as the MI300x and MI355x
Collaboration with research and infrastructure teams to turn systems improvements into real-world model training and inference gains