SV Advanced Computing, LLC
SV Advanced Computing is an independent performance engineering consultancy based in Lafayette, Colorado, specializing in GPU computing, HPC systems, scientific simulations, and AI infrastructure optimization.
Founded by Srinath Vadlamani, Ph.D., the practice draws on 15+ years of experience across national laboratories, semiconductor vendors, and enterprise HPC — from exascale GPU system optimization at DOE/NNSA facilities to novel architecture performance analysis at Arm and HPE — grounded throughout in an understanding of computing as the new experimental arena of science.
Motivation
Behind the profiling and the benchmarks is one driving conviction: computing exists to advance science. Discovery increasingly depends on simulations and models you can actually trust — and that trust has to be earned, against physical ground truth, with reproducible method and honesty about what a number does and does not say. That is why this work is held to a high-fidelity standard rather than a "good enough" one: faster and more trustworthy computation enables faster and more trustworthy science. That conviction took shape across the mathematics of plasma physics, the study of dynamical systems, and exascale scientific applications — and it now carries a real enthusiasm for AI as a way to aid and accelerate that science, with the drive to learn it in all its forms and on all the hardware it runs on. The throughline is a single goal: to make computing a better experimental arena for science.
Principal
Srinath Vadlamani holds a Ph.D. in Applied Mathematics from the University of Colorado Boulder, with a dissertation focus on computational plasma physics and particle-in-cell simulations using dynamical-systems analysis. His career has spanned the full stack of high-performance computing:
- GPU Computing — Deep expertise in both CUDA and ROCm/HIP ecosystems, including kernel optimization, profiling, and performance analysis across NVIDIA and AMD architectures
- Exascale Systems — Performance engineering on the El Capitan exascale system (DOE/NNSA), optimizing scientific workloads on AMD MI300A GPUs
- Parallel Computing — MPI, OpenMP, NCCL/RCCL communication optimization, multi-node scaling analysis
- Scientific Computing — Plasma physics simulation, PDE solvers, numerical methods at scale
Previous affiliations include Tech-X Corporation, the National Center for Atmospheric Research (NCAR), ParaTools Inc., Arm Inc., and Hewlett Packard Enterprise.
Download full CV (PDF) — twenty years across national labs, semiconductor vendors, and enterprise HPC, with the full publication list and toolchain.
Research Interests
Current investigations focus on the intersection of traditional HPC performance engineering and modern AI workloads:
- GPU memory hierarchy analysis and kernel-level optimization
- Custom instrumentation and roofline modeling beyond vendor tools
- Cross-vendor GPU performance comparison (NVIDIA vs. AMD architectures)
- OS and runtime interactions with accelerators — how host-side memory allocation (pinned vs. pageable, NUMA placement, page-fault behavior, unified-memory paging) and allocator pressure gate accelerator throughput, surfacing as stalls a kernel profiler attributes to the GPU but that originate in the operating system and runtime
- Concurrency correctness and contention — thread contention, synchronization-primitive cost, and data races in concurrent host/device code, including how such defects hide at low precision and shift across GPU generations
- System architecture exploration through simulation (gem5, McPAT) — memory-hierarchy and energy-performance trade-offs
- AI efficiency analysis — where inference and training actually spend time and energy: arithmetic intensity, memory-bandwidth vs. compute bounds, batching and utilization, and tokens-per-joule — closing the gap between advertised and delivered performance
Actively Developing
Areas where I'm building hands-on depth through systematic benchmarking and profiling:
- Inference runtime optimization (vLLM, TRT-LLM, SGLang)
- AI training framework performance characterization (PyTorch, NeMo, Megatron-Core)
Contact
For consulting inquiries, collaboration, or questions about our published research and blog posts:
Email: srinath@svac-llc.com
Curriculum vitae: Download CV (PDF)
Colophon: this website is itself the product of a coordinated fleet of Claude Code agents working in parallel — independent sessions, each in its own lane and working directory, kept from colliding by a resident coordinator. That practice is documented in Parallel AI Agents: the system that built this site.