DEEPSEEK Open Source Week Day THREE - DeepGEMM First Look

Video Description

Timestamps: 00:00 - Intro 00:32 - DeepGEMM Overview 07:05 - H100 Test 12:10 - Stock Comparison 14:26 - Neat Tricks 18:57 - Closing Thoughts DeepSeek’s Open Source Week continues with DeepGEMM, a technical deep-dive into FP8 training optimizations for NVIDIA Hopper GPUs. This repository provides insight into how DeepSeek leverages fine-grained scaling techniques to enable faster training and inference while maintaining model quality. In this video, we explore the DeepGEMM repository, breaking down its approach to FP8 training and how it improves efficiency over traditional methods. We install and run the test script on an H100, comparing its performance to a BF16 implementation to analyze the speed and efficiency gains. Along the way, we highlight some of the key tweaks and optimization tricks that make DeepGEMM a powerful tool for accelerating AI workloads. This video provides a first look at DeepSeek’s approach to FP8 training, shedding light on its potential long-term impact on AI model training and inference performance.