Sanjay Suthar

• Designed and developed neural network firmware (control/data path) for Qualcomm Cloud AI 100 PCIe card, orchestrating workloads across Linux runtime, ARM cores, and Hexagon DSP.
• Implemented MultiCard AI firmware to enable tensor-sliced execution across multiple PCIe devices using peer-to-peer (P2P) communication, scaling large language models (GPT-4, Codegen 70 Billion parameters).
• Implemented collective communication library (like NCCL) for Neural Processor Unit firmware, enabling distributed execution of deep learning workloads across multiple compute nodes.
• Designed and implemented a command processor to support PyTorch custom operators and efficiently schedule operators across multiple threads using barrier synchronization.
• Engineered high-performance inter-processor communication (IPC) between ARM cores and Hexagon DSP with cache coherency, achieving multi-GB/s throughput.
• Optimized neural network pipeline to reduce ResNet50 latency from 45 ms to 7 ms and improved LLM tokens/sec performance.
• Integrated Secure Boot with ARM TrustZone and OP-TEE, enabling cryptographically signed and encrypted firmware/model and input loading for confidential computing.
• Performed SoC and board bring-up: DDR calibration, power/clock initialization, device tree setup, and enabling high-speed interfaces (PCIe, SMBus).
• Developed and integrated Linux kernel drivers for PCIe DMA bridge, SPI-NOR flash, MTD, EDAC and error handling subsystems for RAS feature
• Conducted crash dump and memory corruption analysis, resolving DMA timeouts, system hangs, and improving RAS/BMC module robustness.
• Discovered the HW Bugs in Internal IP RTL and PCIe ATU and collaborated with silicon teams to validate the fixes.
• Tools/Technologies: HexagonDSP, PCIe controller, MMU/TLB, ELF loader, RTOS (Qurt), Linux kernel, TrustZone, OP-TEE, JTAG, Trace32, Yocto, CUDA.

Project 1: 3D Audio Framework on i.MX8MM SoC

• Developed a low-latency audio framework with Dolby/DTS/Fraunhofer codec support, HDMI EDID detection, and voice services integration on LK kernel
• Worked on the Jailhouse static hypervisor to partition the NXP IMX8MM to run the RTOS (M4) along side Linux on A72 to provide real time audio processing.
• Developed virtual sound driver layer for Linux to use the Audio services from RTOS.
• Debugged and optimized synchronization issues, reducing audio glitches.

Project 2: ARM-based i.MX SoC Board Bring-Up & Driver Development

• Ported Linux and QNX BSPs (uBoot, kernel, rootfs), configured device trees, enabled Ethernet, USB, SPI, and I2C support.
• Developed drivers for MIPI-CSI cameras, LVDS displays, and audio codecs, validated video pipelines with V4L2 and Gstreamer based applications.
• Reduced boot time by 30% with kernel tuning and peripheral parallelization.
• Built custom Yocto layers, automated BSP builds, and authored DMA scripts for NXP SDMA.
• Worked on linker scripts, MCUXpresso SDK driver porting for Cortex-M4.

Project 3: OP-TEE OS Integration on Custom i.MX Board

• Integrated OP-TEE secure OS with Linux kernel and developed secure TEE applications.
• Implemented ECDSA with SHA256, fixed crypto acceleration bugs, and resolved kernel hangs from misconfigured clocks.
• Implemented HAB/CAAM secure boot with signed and encrypted firmware.