Hynix has initiated mass production of the world's first 321-layer NAND, a triple-level cell-based 4D memory with a capacity of 1Tb, marking a significant advancement in memory technology.
Computing with Time: Microarchitectural Weird Machines
Microarchitectural weird machines ( µ WMs) leverage unobservable events in CPU microarchitecture to create advanced obfuscation techniques, enabling computations that evade traditional detection methods. This approach utilizes side effects from components like branch predictors and caches, allowing malware to operate invisibly until triggered.
A Short Introduction to Automotive Lidar Technology
Lidar technology is essential for level 4 autonomous driving, enabling self-driving taxis like those from Waymo and Cruise, but its widespread adoption hinges on significantly reducing costs, which currently exceed thousands of dollars per unit.
Now Hear This: World’s Most Flexible Sound Machine Debuts
Fugatto, NVIDIA's latest generative AI model, can create and transform music, voices, and sounds from text and audio inputs, showcasing unprecedented flexibility in audio generation.
[R] Aurora: A General-Purpose Foundation Model for Earth System Prediction
Aurora is a foundation model trained on over 1M hours of atmospheric data, enabling it to perform diverse weather and climate predictions with a single architecture, marking a significant evolution in environmental modeling.
Scientists Make Smartphone App That Performs Full-Body Motion Capture
MobilePoser transforms smartphones into movie-quality full-body motion capture systems, utilizing existing sensors in devices like smartphones and smartwatches, making advanced motion capture accessible to all.
[2411.15100] XGrammar: Flexible and Efficient Structured Generation Engine for Large Language Models
XGrammar introduces a flexible and efficient structured generation engine that significantly enhances the performance of large language models (LLMs) by achieving up to 100x speedup in context-free grammar execution.
[R] Testing the Brittleness of LLM Analogical Reasoning Through Problem Variants
Researchers developed a systematic framework to test LLM analogical reasoning, revealing that while models excel at basic transformations, they exhibit a 30-40% drop in performance on multi-step transformations and struggle with novel alphabet systems.
LazyGraphRAG: Setting a new standard for quality and cost
LazyGraphRAG introduces a cost-effective, scalable approach to graph-enabled retrieval-augmented generation (RAG), eliminating the need for prior data summarization and significantly reducing indexing costs to 0.1% of full GraphRAG expenses.
🤗You could have designed state of the art positional encoding
The article details the evolution of positional encoding in transformer models, culminating in Ro tary P ostional E ncoding (RoPE), which enhances self-attention mechanisms by encoding relative positions through rotation rather than addition, improving performance in models like Llama 3.2.
Accelerating drug discovery with TamGen: A generative AI approach to target-aware molecule generation
TamGen, a transformer-based chemical language model, accelerates drug discovery by generating target-specific compounds, achieving significant bioactivity against tuberculosis (TB) protease inhibitors.
The Power of Small Brain Networks
Small brain networks, such as those in fruit flies, can perform complex tasks with as few as four neurons, challenging previous assumptions about the computational capabilities of small neural circuits.
Supercharging Training using float8 and FSDP2
Achieving up to 50% throughput speedup in training with float8 and FSDP2 allows for efficient model training without sacrificing quality, demonstrated across various Meta LLaMa model sizes from 1.8B to 405B.