EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·26d ago·5 min read
Accurate Models of AMD Matrix Cores
TL;DR
Accurate models of AMD's matrix cores were developed to address discrepancies in floating-point computations across different GPU architectures. These models enable reproducibility and facilitate comparisons with NVIDIA's tensor cores.
✦ Why It Matters
Use these models to ensure consistent floating-point computations across AMD GPU architectures in your projects.
Key Takeaways
- Implement MATLAB models for accurate matrix multiplication results.
- Utilize test vectors to identify numerical behavior in GPUs.
- Refine models iteratively for improved accuracy.
- Compare AMD and NVIDIA architectures for application-level accuracy.
Full Summary
Recent AMD GPUs feature matrix multipliers that deviate from the IEEE 754 floating-point standard, leading to inconsistencies in results across different architectures. This study characterizes the numerical behavior of matrix multipliers in AMD's CDNA 1, CDNA 2, and CDNA 3 architectures using MI100, MI210/250, and MI300A/300X GPUs.
Test vectors were designed to probe specific numerical features, and MATLAB-based software models were created and validated against hardware through a randomized test suite of 10 million input vectors. The models were refined iteratively to ensure bit-level reproducibility.
As a demonstration, the models were used to quantify accuracy differences between AMD's matrix cores and NVIDIA's tensor cores, highlighting the practical implications of these discrepancies in real applications.