On-device ML Performance Engineer, Graphics, Games and Machine Learning

The On-Device Machine Learning team at Apple is responsible for enabling the Research to Production lifecycle of cutting edge machine learning models that power magical user experiences on Apples hardware and software platforms. Apple is the best place to do on-device machine learning and this team sits at the heart of that discipline interfacing with research SW engineering HW engineering and products. nnThe On-device ML Performance team has the responsibility to analyze latency memory power and numerical correctness of the latest machine learning models running on Apple SoCs and to make Apples ML software stack take full advantage of the capabilities in Apples ML accelerators. The work from this cross functional team enables model developers decisions to optimize performance via advanced techniques such as different model authoring techniques quantization sparsity performance and accuracy tradeoffs. The work of this team impacts all new Apple HW and ML Inference on group is looking for an On-device ML Performance Engineer with technical expertise in computer architecture performance memory power ML model architectures ML frameworks such as PyTorch and on-device ML inference. The role entails deep analysis of ML models and their architecture the implementation of the models in the ML SW stack for optimum performance power and memory usage and debug involving the performance and power consumption of CPU GPU and Apple Neural Engine.

As an engineer in this role you will be primarily focused on analyzing and optimizing the performance of the latest ML models on the latest iPhones and Macs. You will work with models created by the most popular ML frameworks (PyTorch MLX etc) and will analyze the inference of those models on device to ensure the stack achieves full machine performance on Apple Silicon. The role also includes scripting coding model import and conversions and generation of utilities and debug tools to extract analyze and report performance and power related metrics for Apple HW. The ideal candidate will have a passion for ML model architectures and ML inference deep knowledge of GPU and CPU computer architecture compilers and has a natural inclination toward innovation and exploration.

Driving the on-device performance analysis of Apple SoCs and ML SW stack across a wide range of Apple internal or open-source ML modelsnUsing innovative ways to optimize model conversion compilation and on-device inference for Apple SoCs achieving ambitious goals for performance memory and energy efficiencynDeveloping tools and scripts to generate and analyze ML performance datanWork across multiple teams and organizations to support the design and delivery of best in class on-device ML hardware and software stacknGenerate and present ML performance data to internal and external teams and stakeholders nWork across multiple teams and organizations to support the design and delivery of best in class on-device ML software stack

Experience with ML inference quantization performance and accuracynFamiliarity and experience with the most popular ML architectures (e.g. LLMs Diffusion models CNNs)nA passion to explore and learn about the latest advances in ML model design and architecture particularly as related to model implementation on HW and on-device inferencenFamiliarity with Operating Systems embedded systems and CPU/GPU/SoC/Memory HW architectures nHighly proficient in Python/C and shell scriptingnFamiliarity with Linux or macOSnExceptional clarity in verbal and written communication including the ability to summarize present and lead discussions in larger groups

Masters or PhDs in Computer Science or relevant with Apples CoreML MPS Graph Metal Performance Shaders or MLX frameworksnExperience with any ML authoring framework (PyTorch TensorFlow JAX etc.)nExperience with implementation of high performance compute kernels for CPU GPU or AI AcceleratorsnExperience with Apples App development framework such as Xcode Swift Objective-CnExperience with any on-device ML stack such as TFLite ONNX ExecuTorch with any compiler stack (MLIR/LLVM/TVM etc.)