Experience and education

Total experience: 6 to 10 years in software engineering, mostly systems or infrastructure software (agents, daemons, drivers, platform services), not web or app development.

GPU experience: 3+ years writing software that monitors, manages or tests GPU servers, with at least 1 year hands-on on AMD Instinct and ROCm, for example at a server or GPU vendor, an HPC centre, a GPU cloud or a data-centre infrastructure team.

Hardware you have worked on: any AMD Instinct generation is enough. You do not need to have worked on every model in our fleet; we test each one on real hardware here. Strong NVIDIA GPU systems engineers with real ROCm exposure who want to go deep on AMD are welcome to apply.

Education: Bachelor’s degree in Computer Science, Electronics or a related field, or equivalent practical experience.

This is systems work, not ML work: GPU experience only in model training, PyTorch, CUDA kernels or inference tuning does not meet the bar.

Must-have skills

Go in production for at least 2 years, and C: you have called a C library from Go (or written the C side) in a long-running agent or service.

AMD SMI, ROCm SMI or amdgpu sysfs from your own code, on any AMD Instinct GPU, not only the amd-smi or rocm-smi command line.

Linux systems work on bare metal with the amdgpu driver: PCIe, sysfs and debugfs, kernel logs, driver and ROCm install and versions.

GPU faults on real servers: you have dealt with RAS or ECC errors, bad pages or GPUs that dropped off the bus, and know how to read the evidence.

Metrics design: labels, cardinality, retention and alert rules in Prometheus, VictoriaMetrics or similar.

Key responsibilities

Characterise every AMD model. Test each model and ROCm version, record which metrics, events and controls it actually exposes,



and keep the per-model capability map current.

Build the AMD collector in Go. CGo bindings to the AMD SMI library, plus direct reads of amdgpu sysfs; lightweight and secure on production GPU hosts; tolerant of driver reloads and GPUs that fall off the bus.

Parse gpu_metrics across versions. Read every layout version correctly, and fail safe on a version you have not seen yet.

Handle architecture differences. Temperature mapping by generation, node power mode, and the APU’s shared CPU and GPU memory.

Detect and classify health events. Map RAS blocks, CPER records, bad pages and violation counters onto the platform’s common severity scale, shared with the NVIDIA side, and define where an APU CPU-side error becomes a GPU problem.

Feed metrics into VictoriaMetrics and set thresholds. Metric names, labels and retention that hold at fleet scale, including CPX and NPS partitions; default thresholds per model and per scope.

Build AMD diagnostics. Deep health checks on selected idle GPUs, built on the low-level interfaces, turned into structured, exportable reports.

Implement GPU control safely. Scoped resets, clocks, power limits and CPX/NPS changes with idle checks, read-back, settings re-applied after reboot, batch execution as durable workflows, and audit logging.

Support field issues. Reproduce and root-cause customer GPU issues, and feed the lessons back into detection rules, thresholds and documentation.

Pay: ₹1,500,000.00 - ₹1,570,000.00 per year

Benefits:

- Paid sick time
- Paid time off
- Provident Fund
- Work from home

Application Question(s):

- What is your current CTC, Expected CTC & Notice Period Time?

Experience:

- software engineering: 6 years (Required)
- GPU : 3 years (Required)
- Go in production : 2 years (Required)

Work Location: Remote

📌 Senior GPU Infrastructure Engineer, AMD (India)
🏢 SecHorizon Technologies
📍 India