High Performance Computing

My way into HPC began as plain curiosity about how you actually program a GPU, the same kind of chip that runs the video games I grew up on. That question led me to supercomputers, which are racks of the same silicon doing work far heavier than rendering frames. Half a decade in, what still excites me most is the chase, hunting performance through layers of code and silicon until I find the one function that's quietly bottlenecking the entire pipeline. A bottleneck is the one slow step that everything behind it has to wait for, like the neck of a bottle setting the pace for everything that pours through it, and a program can never run faster than that step.

What I love is the layering of the work. Chasing performance bottlenecks sits alongside keeping a cluster healthy and redesigning a CI pipeline that touches thirty applications. A cluster is a set of computers wired together and run as one machine, which is what those racks are. CI is continuous integration, the practice of having a machine build and test every change to a piece of software as soon as it's made, so that a mistake shows up the same day and not at release. The pipeline is the chain of automated steps that does it. Learning to use these machines well, and helping others do the same, is what I plan to spend my career on.

A supercomputer is how you get an answer that would otherwise take a lifetime. That is why a five-day weather forecast is now as good as a one-day forecast was about forty years ago, and why a molecule can be tried on a machine before anyone makes it in a lab. The chips stopped getting faster on their own a while back, so from here on most of the speed has to come out of the software. The old rule of the trade still holds - the part of a program you leave alone is the part that decides how fast the whole thing can ever go. The Gita's line about having a right to the work and never to its fruits fits this job better than most, which is the frame of mind the hunt for that one function needs.

What follows is that journey so far - the roles I've held, the awards, my education, the research, and some writing.

Work Experience

AMD

Current

Full-time

Aug 2021 – Present · 5 yrs 2 mos

Bangalore, IN

Sr. Software System Designer

Dec 2025 – Present · 10 mos

Recently promoted - highlights to follow.

Software System Designer 2

May 2023 – Nov 2025 · 2 yrs 7 mos

I owned three applications in every release of Zen Software Studio from 4.2 to 5.2, AMD's bundle of compilers, maths libraries and profiling tools for its Zen processors. The three were GROMACS, which simulates how molecules move, HMMER, which searches protein and DNA sequences for their relatives, and SPEC MPI, a standard set of programs for measuring how fast a cluster is. Owning them meant getting each one to build through Spack and through Kaveri, sorting out any performance problem that came up, and coordinating the public documentation for building them with Spack. Spack is the package manager supercomputing centres use to install scientific software with everything it depends on.
I designed and built the team's CI platform on Jenkins, an open source server that runs builds and tests on its own, and on Kaveri. It automates more than thirty HPC applications across more than ten toolchains and more than eighty workloads. A toolchain is the compiler and libraries a program is built with, and a workload is the input an application is run on. It took a large part of the effort out of every release cycle. Spotlight Award, Q4 2024.
I led ZSS Operations for the 5.0, 5.1 and 5.2 releases, which meant running the release work itself and reporting the performance numbers, and I brought two contractors on board so the work could grow.
GROMACS had a 4 to 5% performance gap against Intel's oneAPI, Intel's own toolkit of compilers and libraries. I closed it by digging down to the root cause with the AOCC team, AOCC being AMD's compiler for C, C++ and Fortran, and with the Performance Analysis team.
I led the adoption of Spack across the organization. That meant writing the recipes, the instructions Spack follows to build a package, for our HPC applications, and getting them running on the MI300, AMD's accelerator for data centres, so that it could be validated. It also meant sending our internal recipes upstream to the public Spack project, so that anyone could use them. Spotlight Award, Q2 2023.
I automated the way we make and submit changes to Spack recipes, hence contributing them upstream became routine. Director Spotlight Award, Q1 2025.
I automated the deployment of the Dhruva1 cluster with Ansible, a tool that sets up any number of machines from one written description instead of by hand. Spotlight Award, Q3 2023.
I started a collaboration between AMD and EPCC, the University of Edinburgh's supercomputing centre, which included a joint proposal for a dissertation and widened AMD's presence in academic HPC.

Software System Designer 1

Aug 2021 – Apr 2023 · 1 yr 9 mos

I measured what AVX-512 does for GROMACS and profiled it on the fourth generation of EPYC, AMD's server processors. AVX-512 is a set of processor instructions that work on many numbers in one go, and profiling means recording where a program spends its time. What I found went into Zen Software Studio's optimizations for those chips.
I wrote a tool that packs an HPC application into a container on its own, and it cut the effort of deploying an application by over 70%. A container is the program packed up with everything it needs to run, so that it runs the same on any machine. Spotlight Award, Q3 2022.
I built an HPC cluster from scratch and looked after the team's hardware and software.
I built a proof of concept on OpenStack to bring the organization's hardware for CI, for delivering builds and for performance validation under one roof. OpenStack is software that turns a room of servers into a private cloud that hands out machines on request.
I wrote and maintained the first Spack recipes for our HPC applications, which is the groundwork the adoption across the organization was later built on.

NVIDIA

Internship

Jan 2020 – Jul 2021 · 1 yr 7 mos

Bangalore, IN

HPC/AI Research Intern

I measured how DFT-FE, a code that works out the electronic structure of a material from first principles, performs and scales on a DGX A100 cluster, and saw it run twice as fast as on the older DGX V100 cluster. Scaling is how much faster a code goes as it is given more GPUs, and DGX is NVIDIA's GPU server, with A100 and V100 two generations of the GPU inside it.
With the HPC team I helped port the CFS code onto GPUs with OpenACC, a way of marking up existing code so that the compiler moves the marked parts onto a GPU, and profiled the result.
I researched and built a model that generates Indian classical music, using MuseGAN, a neural network that composes music as several instrument tracks at once.
I wrote and profiled several machine learning algorithms on GPUs for RAPIDS, NVIDIA's data science libraries that run on the GPU, and for SimNet, its toolkit for training neural networks that obey the laws of physics.

Awards & Recognition

Hackathons
· AMD
1 award · 2026
- Winner, CSG AI Hackathon 2026 Feb 2026
  
  For the entry "From Few Experts to Many: AI-Democratized Kaveri Expertise", which did what its title says, using AI to put what a few Kaveri experts know within everyone's reach. A hackathon is a day or two in which teams build something that works from nothing.
Spotlight Awards
· AMD
6 awards · 2022–2026
- Director Spotlight Award, Q2 2026 Jun 2026
  
  For building Kaveri's web front end, the part of it people see and click, against a short deadline.
- Director Spotlight Award, Q1 2025 Mar 2025
  
  For the automation of changes to Spack recipes.
- Spotlight Award, Q4 2024 Dec 2024
  
  For designing and building the HPC team's CI platform on Jenkins and Kaveri.
- Spotlight Award, Q3 2023 Sep 2023
  
  For automating the deployment of the Dhruva1 cluster with Ansible.
- Spotlight Award, Q2 2023 Jun 2023
  
  For the Spack recipes that let the HPC applications be validated on the MI300.
- Spotlight Award, Q3 2022 Sep 2022
  
  For making the case for running HPC applications in Docker containers for benchmarking.
Director of Engineering Appreciation
· NVIDIA
Jul 2021

A note of appreciation from the Director of Engineering for my work during the internship.

Education

University of Edinburgh
Sep 2023 – Sep 2026

M.Sc. · High Performance Computing

High Performance Computing at EPCC, the supercomputing centre of the university, studied online from Bangalore alongside the job at AMD. The dissertation, on what hardware counters can and cannot tell you about a program, is under Research below.
PES University
Aug 2017 – Aug 2021

B.Tech. · Computer Science and Engineering

Computer Science and Engineering at PES University in Bangalore. A B.Tech. is India's four-year engineering degree, and the paper on access control under Research came out of this time.

Research

Aug 2026 Dissertation HPC Performance 1 hr 39 min read

What hardware counters can and cannot tell us

MSc Dissertation, EPCC, University of Edinburgh · Edinburgh, UK

I spent a year testing whether CPU performance counters can predict how fast an HPC application will run. They can't. They explain runtime rather than project it, and the useful part of the project was measuring exactly where that boundary sits and what crossing it costs.

May 2021 Paper Security IoT

Access Control for Streaming Data Across Heterogeneous Communication Models

International Conference on Artificial Intelligence for Internet of Things · Seattle, US

A unified mechanism to insert access-control policies into data streams and enforce them across heterogeneous communication models.

Oct 2020 Paper Graph Algorithms OpenCL

Heterogeneous Parallel Critical Nodes Detection

International Conference on Advanced Computer Science and Information Systems · Depok, ID

A heterogeneous algorithm for the Critical Node Detection Problem using OpenCL, outperforming the earlier parallel implementation by 2× on real-world Twitter data.

Writing

Jun 2026 Technology 16 min read

Ten models, four domains

I built a local AI coaching system for Hindustani music, tennis, photography and career coaching, then spent two weeks benchmarking ten models to find the right one, on speed, on quality, on whether letting a model think first helps, and through a lot of out-of-memory crashes.