Deep Learning

Gemini's Inner Workings: A Deep Dive into Tensor Processing and Model Parallelism

The advent of large language models (LLMs) like Google’s Gemini has revolutionized what’s possible in AI. While much attention is paid to their impressive capabilities, the underlying infrastructure and architectural patterns that enable them are a testament to cutting-edge engineering. This post dives deep into the “under-the-hood” aspects of Gemini, focusing on the specialized hardware for tensor processing and the intricate strategies employed for model parallelism. The Tensor Processing Engine: Beyond the CPU At the heart of any LLM’s ability to perform complex calculations lies its proficiency in handling tensors.

Continue reading

Deep-Dive into DeepSeek's MoE Architecture for LLMs

Unpacking DeepSeek: The Power of Sparse Mixture-of-Experts in Large Language Models The landscape of Large Language Models (LLMs) is continuously evolving, with new architectures and training methodologies pushing the boundaries of what’s possible. Among the rising stars in the open-source community, DeepSeek models have garnered significant attention, particularly for their innovative application of the Mixture-of-Experts (MoE) architecture. This deep dive will go beyond mere performance metrics, dissecting the ‘under-the-hood’ mechanisms that make DeepSeek models both powerful and efficient.

Continue reading