Nikita Astafev
Principal Engineer at NVIDIA
Portland, Oregon, United States
Summary
👤
Senior🎓
Top SchoolNikita Astafev is a Principal Engineer based in Portland with 8 years of recent industry experience and a long career designing and optimizing low-level math libraries for CPUs and GPUs. He specializes in floating-point mathematics, HW/SW co-design, SIMD and assembly fine-tuning, and has driven early architecture enablement through compiler and performance-library work at Intel and NVIDIA. His contributions span highly optimized implementations of transcendental functions, RNGs, and math APIs, plus oversight of library engineering and training of customer-facing consultants. Active in open-source performance work (e.g., oneDNN pooling and backward-pass fixes), he combines deep numerical-analysis research—automated FP error analysis and function generation—with practical fixes like offset and threading corrections. Known for finding compiler and simulator deficiencies and for filing multiple patents, he blends rigorous mathematical training from MSU with hands-on systems engineering to accelerate new architectures.
8 years of coding experience
20 years of employment as a software developer
Lomonosov Moscow State University
English, Russian
Github Skills (10)
c-language10
deep-learning10
cpp10
pooling10
cprogramming-language10
batch-normalization10
performance-optimization10
avx9
64-bit8
x86-648
Programming languages (1)
Github contributions (1)
uxlfoundation/oneDNN
May 2018 - Aug 2018
oneAPI Deep Neural Network Library (oneDNN)
Role in this project:
Back-end Developer & Performance Engineer
Back-end Developer & Performance EngineerContributions:21 commits, 2 comments in 2 months
Contributions summary:Nikita primarily focused on optimizing the performance of the oneDNN library, specifically targeting pooling operations for the NHWC format. They implemented and fixed offset computations, and addressed stack corruption issues within the benchmarking framework. Additional contributions included fixes related to large size handling and the correction of inconsistencies in spatial threading decisions for batch normalization. Furthermore, the user worked on backward pass execution for multiple components.
bfloat16sse41avx512openmpopencl
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.