With the increasing growth of artificial intelligence—introduction of large language models (LLMs) and generative AI—there has been a growing demand for more efficient graphics processing units (GPUs). GPUs are specialized hardware extensively used for high computing tasks and capable of executing computations in parallel. Writing proper GPU kernels is important to utilize GPUs to their… →

Liquid AI has released its first series of Liquid Foundation Models (LFMs), ushering in a new generation of generative AI models. These models are positioned as a new benchmark for performance and efficiency at multiple scales, namely the 1B, 3B, and 40B parameter configurations. This series aims to set a new standard for generative AI… →

Multimodal models aim to create systems that can seamlessly integrate and utilize multiple modalities to provide a comprehensive understanding of the given data. Such systems aim to replicate human-like perception and cognition by processing complex multimodal interactions. By leveraging these capabilities, multimodal models are paving the way for more sophisticated AI systems that can perform… →

Microsoft unveiled VoiceRAG, a voice-based retrieval-augmented generation (RAG) system that utilizes the new Azure OpenAI gpt-4o-realtime-preview model to combine audio input and output with powerful data retrieval capabilities. This innovative system represents a significant leap in natural language processing by enabling seamless interaction with applications using voice commands. VoiceRAG is designed to provide a more… →

Traffic forecasting is a fundamental aspect of smart city management, essential for improving transportation planning and resource allocation. With the rapid advancement of deep learning, complex spatiotemporal patterns in traffic data can now be effectively modeled. However, real-world applications present unique challenges due to the large-scale nature of these systems, which typically encompass thousands of… →

Transformer-based Models in Segmentation tasks have initiated a new transformation in the Computer Vision realm. Meta’s Segment Anything Model has proven to be a benchmark due to its robust and exquisite performance. SAM has proven highly successful as supervised segmentation continues to gain popularity in fields such as medicine, defense, and industry. However, it still… →

Large language models (LLMs) have gained widespread popularity, but their token generation process is computationally expensive due to the self-attention mechanism. This mechanism requires attending to all previous tokens, leading to substantial computational costs. Although caching key-value (KV) states across layers during autoregressive decoding is now a common approach, it still involves loading the KV… →