Staff Compiler Engineer - PyTorch + Kernel DSLPLATE

Samsung Semiconductor•San Jose, CA

70d•$163,000 - $253,000•Onsite

About The Position

The AGI (Artificial General Intelligence) Computing Lab is dedicated to solving the complex system-level challenges posed by the growing demands of future AI/ML workloads. Our team is committed to designing and developing scalable platforms that can effectively handle the computational and memory requirements of these workloads while minimizing energy consumption and maximizing performance. To achieve this goal, we collaborate closely with both hardware and software engineers to identify and address the unique challenges posed by AI/ML workloads and to explore new computing abstractions that can provide a better balance between the hardware and software components of our systems. Additionally, we continuously conduct research and development in emerging technologies and trends across memory, computing, interconnect, and AI/ML, ensuring that our platforms are always equipped to handle the most demanding workloads of the future. By working together as a dedicated and passionate team, we aim to revolutionize the way AI/ML applications are deployed and executed, ultimately contributing to the advancement of AGI in an affordable and sustainable manner. Join us in our passion to shape the future of computing!

Requirements

Bachelor’s with 10+ years, or Master’s with 8+ years, or PhD's with 5+ years of industry experience.
3-5+ years of industry experience in at least one of: Triton, Helion, MLIR, XLA, TVM, Inductor, IREE, CUTLASS, or a proprietary equivalent (More experienced candidates will also be considered at relevant levels).
Experience designing a kernel DSL or its IR from scratch, or making non-trivial language-level changes to an existing one.
Experience with MLIR — writing dialects, passes, or backend integration.
Experience building PyTorch backends for non-CUDA accelerators (XPU, ROCm, MPS, TPU, custom).
Experience with kernel autotuning, performance modeling, or cost-based compilation
Background in HPC, distributed systems, or NUMA-aware programming — anything that built intuition for non-flat memory

Nice To Haves

Open-source contributions to PyTorch, Triton, Helion, LLVM/MLIR, or similar projects is a big plus.

Responsibilities

Adapting torch.compile to our backend: lowering Inductor's IR to our hardware, defining what gets fused, what gets specialized, and where the compiler should yield to hand-written kernels.
Building or extending kernel DSLs for our hardware: taking a tile-based programming model (Triton-style), a higher-level expression (Helion-style), or a custom DSL we design, and lowering it to our ISA, our memory hierarchy, and our collective primitives. Where existing DSLs' GPU assumptions break, deciding what to change in the frontend, the IR, or the backend.
Designing placement and scheduling passes: given a graph and our distributed memory model, deciding where tensors live, when to migrate them, and how to overlap compute with data movement. This is the layer where our hardware's differentiator shows up most directly.
Implementing parallelism-aware lowering: making tensor, pipeline, expert, and sequence parallelism first-class in the compiler IR rather than bolted on at the framework layer.
Fusion, tiling, and memory planning: the classical compiler problems, reframed for a non-uniform memory hierarchy where the right tile size and the right placement are coupled decisions.
Upstream contributions: where we use open-source DSLs, we want our work to land upstream rather than live in a private fork. You'll engage with upstream review processes for PyTorch, Triton, Helion, and adjacent projects.