# Accelerating Density Fitting with Adaptive-precision and 8-bit Integer on AI Accelerators

**Authors:** Hua Huang, Wenkai Shao, Jeff Hammond  
**arXiv:** [2601.08077](https://arxiv.org/abs/2601.08077)  
**Date:** 2026-01-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2601.08077v3/

## Key points

- Research explores the potential of AI accelerators in enhancing tensor-heavy scientific computations
- Focus on applying NVIDIA Tensor Cores to quantum chemistry for accuracy and complex data patterns
- Introduces adaptive precision algorithm for accelerating density fitting method using Gaussian basis sets on AI accelerators through 8-bit integer arithmetics
- Algorithm implemented in GPU-accelerated PySCF package and tested across more than twenty molecular systems with various NVIDIA GPUs
- Results show significant performance enhancements compared to conventional FP64 code implementations, with up to 204% acceleration on an RTX 4090 gaming GPU and up to 364% speedup on an RTX 6000 Ada workstation GPU while maintaining converged energy levels
- Demonstrates practical approach for leveraging AI hardware in quantum chemistry simulations and potential for further advancements in accelerating scientific computations through adaptive precision algorithms and optimized utilization of AI accelerators

## Summary

In their research titled "Accelerating Density Fitting with Adaptive-precision and 8-bit Integer on AI Accelerators," authors Hua Huang, Wenkai Shao, and Jeff Hammond explore the potential of artificial intelligence (AI) accelerators in enhancing tensor-heavy scientific computations. Specifically utilizing NVIDIA Tensor Cores, the focus of their study lies in applying these accelerators to quantum chemistry - a field known for its stringent accuracy requirements and complex data patterns. To address these challenges, the researchers introduce an innovative adaptive precision algorithm designed to expedite the density fitting (DF) method using Gaussian basis sets on AI accelerators through 8-bit integer (INT8) arithmetics. Implemented within the GPU-accelerated PySCF package, the proposed algorithm undergoes rigorous testing across more than twenty molecular systems employing various NVIDIA GPUs. The results showcase significant performance enhancements compared to conventional FP64 code implementations. Notably, the algorithm demonstrates up to a remarkable 204% acceleration on an RTX 4090 gaming GPU and up to an impressive 364% speedup on an RTX 6000 Ada workstation GPU while maintaining the requisite level of converged energy in calculations. This groundbreaking work not only highlights a practical approach towards leveraging AI hardware for dependable quantum chemistry simulations but also underscores the potential for further advancements in accelerating scientific computations through adaptive precision algorithms and optimized utilization of AI accelerators.