AIML Pipeline/ Dataflow Design Lead

AIML Pipeline / Dataflow Design Lead

Location: Bangalore, India |Company: LTSCT — L&T Semiconductor Technologies Limited

Reporting To: Senior Architect, DC HW Systems | Experience: 10+ years (mandatory)

Role Summary

We are hiring two AIML Pipeline / Dataflow Design Leads to implement the BFOS-MPE (Broadcast-Fed Output-Stationary Matrix Processing Engine) — the compute heart of xPU. You will own the compute datapath, the broadcast network, PE array control, sparsity engines (SWS), and dual-mode (matrix + GPU) switching. You will also own the LLVM interface to the xPU.

Key Responsibilities

• Own micro-architecture and RTL of the BFOS-MPE compute datapath — dual 256×256 PE arrays (65,536 PEs each) operating at 2 GHz.
• Design the broadcast-fed input network and output-stationary accumulation scheme that define the BFOS dataflow.
• Implement PE array control, sequencing, and the sparsity engine (SWS) for structured/unstructured sparsity acceleration.
• Architect dual-mode operation enabling switching between dense matrix (GEMM/convolution) and GPU-style workloads.
• Drive numerical formats and precision (FP/BF/INT, mixed precision) and validate functional correctness against reference models.
• Coordinate and integrate contributions into the dataflow codebase and verification plan.
• Own performance/utilization modeling, power-efficiency optimization, and PPA closure for the compute engine.
• Own the ISA/FW routines for the optimal utilization of the HW compute structures in the xPU; as part of the LLVM development.

Required Skills & Experience

• 10+ years in compute datapath, DSP, GPU, or AI-accelerator micro-architecture and RTL design.
• Deep understanding of systolic/spatial arrays, matrix-multiply dataflows (output/weight/row-stationary), and broadcast networks.
• Strong expertise in RTL (SystemVerilog) for high-throughput arithmetic datapaths and control.
• Solid grounding in numerical formats and arithmetic (FP32/BF16/FP8/INT8), mixed-precision, and rounding/accuracy trade-offs.
• Experience with sparsity acceleration techniques (structured/unstructured, weight/activation sparsity).
• Familiarity with deep-learning operators (GEMM, convolution, attention) and how they map to hardware.
• Ability to build and correlate performance/utilization models against RTL and reference software.
• Hands-on PPA optimization for large, dense compute blocks.
• Proficiency in C/C++/Python for modeling and verification support.
• Proven leadership and experience integrating cross-organization/partner engineering contributions.
• AI compiler development experience is a definite plus.

Preferred Qualifications

• Direct experience on a taped-out AI/ML accelerator or tensor/matrix engine.
• Familiarity with GPU SIMT execution models and matrix+vector dual-mode designs.
• Exposure to ML compiler/graph mapping (MLIR, TVM) and how software drives the datapath.
• M.Tech/MS/PhD in EE/ECE/CS or equivalent.