Ashwin Nanjappa

Manage and lead teams for: - SemiAnalysis benchmark - Day-0 launch (Llama4) - MLPerf Inference benchmark - TensorRT LLM performance - TensorRT performance

Implemented new DL layers, performance optimizations, performance infrastructure and new GPU architecture bringup for the TensorRT inference library.

I represent NVIDIA at the Working Group (WG) of MLPerf Inference and its benchmark taskforces. Core member of WG, intimately involved in creation of all new benchmarks and rules for MLPerf Inference.

As a core member of the Algorithms team, I created CV algorithms (Matlab, C++), trained DL models (Caffe), developed DL training pipelines (Python) and deployed inference pipelines in the cloud for some of the world's largest ecommerce vendors. Team lead on developing a new training pipeline built around Caffe and Python to train new network structures on multi-GPU systems.

Created and implemented new GPU-accelerated machine vision algorithms for depth camera ► Created: GHand, a realtime GPU algorithm for hand pose estimation from depth camera » Wrote: Entire codebase of C++, CUDA, OpenGL and OpenCV to model, train and test hand estimation » Built: 3D-CUBE augmented reality system for 3D interaction with biomolecules using hand gestures (see video) » Built: HPES, a hand-pose depth image dataset for use by all machine vision researchers » Published in Eurographics and IJCV. Demoed at Singapore Science Festival 2014 and 2015 to public. ► Created: World’s first realtime algorithm for full-body pose estimation of mouse from depth camera » Wrote: Entire codebase of C++, CUDA and OpenGL to model, train and test mouse estimation » Built: Novel two-camera multi-view cage setup that can be used by neuroscientists » My system is being used by neuroscientists at the National Neuroscience Institute

Conducted instruction, tutorials and labs for students at School of Computing, NUS. ► Courses taught: Software Engineering, Computer Graphics and C Programming ► Award: Nominated by my students for Best Teaching. Testimonials and grades from my students available on request.

● Project: Digital Delaunay triangulation in 3D ● Collaborators: Dr. Huang Zhiyong and Dr. Ishtiaq Rasool Khan ● Developed methods to color and dualize a 3D digital Voronoi diagram so that a topologically and geometrically valid triangulation can be dualized from it. Results of this project are part of my PhD thesis.

► Built: C API and code for IP Multimedia System (IMS), Session Description Protocol (SDP) and Session Interface Protocol (SIP) on Motorola cellphones ► Shipped: In millions of cellphones sold on Verizon and other US cellphone networks

► Built: C++ code for SACD and DVD-Audio features on Yamaha and Philips DVD players. ► Shipped: In millions of economy DVD players sold in India and highend players sold in USA/Japan

► My thesis: Delaunay Triangulation in 3D on the GPU ► I created: The fastest GPU algorithms for 3D Delaunay and regular triangulations » gDel3D is a hybrid GPU-CPU 3D Delaunay algorithm. 10x faster than CGAL. » gStar4D is a full-GPU 3D Delaunay algorithm. 5x faster than CGAL. » gReg3D is a full-GPU 3D regular triangulation algorithm. 4x faster than CGAL. » Open source: Full source code of my highly optimized and robust CUDA implementation is available on Github.

► Graduated summa cum laude from R. V. College of Engineering, Bangalore ► Among the top 3 students in the university honor roll for Computer Science in 2002 ► Thesis: 3D Surface Modeling of Serially Acquired Medical Images at Siemens Medical, Bangalore