DTMA & Memory Controller Design Lead

L&T Semiconductor Technologies; Bengaluru; On-site; INR 3,500,000 - 9,000,000; Full time

Job summary

L&T Semiconductor Technologies in Bengaluru, Karnataka is seeking a DTMA & Memory Controller Design Lead to own the DTMA engine and memory-subsystem integration. You will orchestrate data movement between large SRAM banks, HBM4e stacks, and PE arrays to keep the compute engine saturated.

The role requires 10+ years in memory-subsystem design, strong SystemVerilog RTL skills, and hands-on modeling in C/C++/Python. Collaboration with PD and architecture on memory hierarchy is expected.

Qualifications

• 10+ years in memory-subsystem, DMA, or memory-controller design for high-bandwidth SoCs.
• Deep expertise with DRAM/HBM controller architecture, scheduling, and timing (HBM2E/HBM3/HBM3E; HBM4e a strong plus).
• Strong RTL (SystemVerilog) design skills for DMA engines, arbiters, and memory schedulers.
• Solid understanding of multi-bank SRAM architectures, banking/interleaving, and conflict resolution.
• Experience with descriptor-based DMA, gather/scatter, tensor tiling, and prefetching.
• Knowledge of address mapping, request reordering, and bandwidth/efficiency optimization.
• Familiarity with AI/ML data-movement patterns and their impact on memory-system design.
• Hands-on performance modeling and closure for high-bandwidth memory subsystems.
• Proficiency in C/C++/Python for modeling and verification.
• Proven technical leadership and cross-functional collaboration.

Responsibilities

• Own micro-architecture and RTL of the DTMA engine managing data movement between SRAM banks, HBM4e, and the BFOS PE arrays.
• Define the compute-to-memory interface with the DTMA engine and SRAM banks to keep PE arrays fully fed at bandwidth.
• Lead integration and configuration of HBM4e memory controllers across 6 stacks (384 GB, 12 TB/s aggregate).
• Design descriptor-based DMA, gather/scatter, tensor tiling, and prefetch mechanisms optimized for BFOS dataflow.
• Architect the 512-bank, 8 TB/s SRAM access scheduling, bank arbitration, and conflict-avoidance logic.
• Optimize address mapping, request scheduling, and refresh/timing to maximize sustained HBM bandwidth efficiency.
• Define QoS and flow-control interfaces to the NoC and compute engine to prevent stalls and starvation.
• Drive verification, performance modeling, and bandwidth/latency closure across realistic AI workload traffic.
• Collaborate with PD on HBM PHY/controller placement and with architecture on memory-hierarchy trade-offs.

Skills

Memory subsystem design

DMA engines

SystemVerilog RTL

HBM controller

C/C++/Python modeling

High bandwidth memory

Job description

Role Summary

The DTMA & Memory Controller Design Lead owns the Data Transfer and Memory Access (DTMA) engine together with HBM4e memory-controller integration. You will orchestrate all data movement between the 512-bank SRAM (8 TB/s), the 6-stack HBM4e subsystem (384 GB, 12 TB/s), and the PE arrays — the subsystem that keeps xPU's compute engine saturated.

Key Responsibilities

• Own micro-architecture and RTL of the DTMA engine managing data movement between SRAM banks, HBM4e, and the BFOS PE arrays.
• Define the compute-to-memory interface with the DTMA engine and SRAM banks to keep PE arrays fully fed at bandwidth.
• Lead integration and configuration of HBM4e memory controllers across 6 stacks (384 GB, 12 TB/s aggregate).
• Design descriptor-based DMA, gather/scatter, tensor-tiling, and prefetch mechanisms optimized for BFOS dataflow.
• Architect the 512-bank, 8 TB/s SRAM access scheduling, bank arbitration, and conflict-avoidance logic.
• Optimize address mapping, request scheduling, and refresh/timing to maximize sustained HBM bandwidth efficiency.
• Define QoS and flow-control interfaces to the NoC and compute engine to prevent stalls and starvation.
• Drive verification, performance modeling, and bandwidth/latency closure across realistic AI workload traffic.
• Collaborate with PD on HBM PHY/controller placement and with architecture on memory-hierarchy trade-offs.

Required Skills & Experience

• 10+ years in memory-subsystem, DMA, or memory-controller design for high-bandwidth SoCs.
• Deep expertise with DRAM/HBM controller architecture, scheduling, and timing (HBM2E/HBM3/HBM3E; HBM4e a strong plus).
• Strong RTL (SystemVerilog) design skills for DMA engines, arbiters, and memory schedulers.
• Solid understanding of multi-bank SRAM architectures, banking/interleaving, and conflict resolution.
• Experience with descriptor-based DMA, gather/scatter, tensor tiling, and prefetching.
• Knowledge of address mapping, request reordering, and bandwidth/efficiency optimization.
• Familiarity with AI/ML data-movement patterns and their impact on memory-system design.
• Hands-on performance modeling and closure for high-bandwidth memory subsystems.
• Proficiency in C/C++/Python for modeling and verification.
• Proven technical leadership and cross-functional collaboration.

Preferred Qualifications

• Experience integrating HBM PHY/controller IP on advanced nodes.
• Background in AI accelerator or HPC memory subsystems.
• Familiarity with UCIe/NoC traffic interaction with memory controllers.