NVIDIA introduces AI-powered fusion kernels for MoE training throughput

NVIDIA has introduced new AI-powered fusion kernels to boost Mixture-of-experts (MoE) training throughput. These kernels deliver a 1.3x-2x kernel-level speedup and have contributed to performance improvements in NVIDIA's DeepSeek-V3 and GPT-OSS pre-training setups.

Researchers have developed Flash-KMeans, an IO-aware, exact K-Means library that runs over 200x faster than FAISS on GPUs. This library targets modern AI pipelines and achieves up to 17.9x end-to-end speedup over optimized baselines.

Kaspersky highlights that AI agents will not replace attackers but will scale them, giving them more power and capacity. The company emphasizes a practical approach to cyber threats and stresses the importance of fostering a healthy security culture within organizations.

Courts are cracking down on error-strewn AI-assisted legal briefs, and lawyers are pushing back against the trend, arguing for more guidelines on AI usage.

Meta is working on a smart glasses app that uses face recognition, partnering with Rank One, a Pentagon supplier. The app aims to identify people and provide personalized information and services.

Thales has unveiled a new AI-powered training data analytics platform to enhance military training, transforming raw data from training exercises into structured intelligence for after-action reviews.

Steam's Next Fest event saw a surge in AI-generated games, with many developers using AI tools to create demos. Some smaller teams and solo devs used AI as an excuse for lacking resources.

The AI era has increased cybersecurity risks, with AI-enabled workflows exposing internal data to cyberattacks. A zero-trust architecture approach can help lower data-breach costs.

Researchers have built an AI-driven worm that adapts in real-time, spreading by commandeering victims' computing power and stealing data.

Gartner forecasts that AI will drive a 26% surge in data center power consumption to 565TWh this year, making data center power security a critical concern.

Key Takeaways

['NVIDIA introduced new AI-powered fusion kernels to boost Mixture-of-experts (MoE) training throughput.', 'Flash-KMeans, an IO-aware, exact K-Means library, runs over 200x faster than FAISS on GPUs.', 'Kaspersky emphasizes that AI agents will not replace attackers but will scale them, giving them more power and capacity.', 'Courts are cracking down on error-strewn AI-assisted legal briefs.', 'Meta is working on a smart glasses app that uses face recognition.', 'Thales unveiled a new AI-powered training data analytics platform to enhance military training.', "Steam's Next Fest event saw a surge in AI-generated games.", 'The AI era has increased cybersecurity risks, with AI-enabled workflows exposing internal data to cyberattacks.', "Researchers built an AI-driven worm that adapts in real-time, spreading by commandeering victims' computing power.", 'Gartner forecasts a 26% surge in data center power consumption to 565TWh this year driven by AI.']

NVIDIA boosts AI training with fusion kernels

NVIDIA has introduced new AI-powered fusion kernels to boost Mixture-of-experts (MoE) training throughput. These kernels, built with the CuTe DSL, tackle memory and synchronization bottlenecks, delivering a 1.3x-2x kernel-level speedup. In NVIDIA's DeepSeek-V3 pre-training setup, this optimization contributes an 8% end-to-end performance improvement. For the GPT-OSS pre-training setup, it contributes a 93% end-to-end performance improvement.

Flash-KMeans speeds up data clustering

Researchers have developed Flash-KMeans, an IO-aware, exact K-Means library that runs over 200x faster than FAISS on GPUs. It targets modern AI pipelines where k-means is called inside training and inference loops. Flash-KMeans restructures the algorithm to move data efficiently on a GPU, achieving up to 17.9x end-to-end speedup over optimized baselines.

AI's impact on cybersecurity

Kaspersky highlights that AI agents will not replace attackers but will scale them, giving them more power and capacity. The company emphasizes a practical approach to cyber threats, noting that human inventiveness continues to dominate successful compromises. Kaspersky stresses the importance of fostering a healthy security culture within organizations.

AI-assisted legal briefs face scrutiny

Courts are cracking down on error-strewn AI-assisted legal briefs. The use of AI in legal drafting is rising, but it can lead to errors and undermine the integrity of the legal system. Lawyers are pushing back against the trend, arguing for more guidelines on AI usage.

Meta tests face recognition for smart glasses

Meta is working on a smart glasses app that uses face recognition, partnering with Rank One, a Pentagon supplier. The app aims to identify people and provide personalized information and services. The technology is still in the prototype phase.

Thales enhances military training with AI

Thales has unveiled a new AI-powered training data analytics platform to enhance military training. The platform transforms raw data from training exercises into structured intelligence for after-action reviews. It provides a comprehensive view of field events, enabling informed decisions and improved operational performance.

Steam flooded with AI games

Steam's Next Fest event saw a surge in AI-generated games, with many developers using AI tools to create demos. Some smaller teams and solo devs used AI as an excuse for lacking resources, while others claimed AI was used for minor tasks.

Cybersecurity in the AI era

The AI era has increased cybersecurity risks, with AI-enabled workflows exposing internal data to cyberattacks. A zero-trust architecture approach can help lower data-breach costs. Data-centric controls, such as encryption, are needed to protect sensitive information.

AI worm adapts in real-time

Researchers have built an AI-driven worm that adapts in real-time, spreading by commandeering victims' computing power. The worm clones itself across devices, stealing data and bypassing weak passwords.

AI drives data center power surge

Gartner forecasts that AI will drive a 26% surge in data center power consumption to 565TWh this year. The growth in AI deployments is straining global grid capacity, making data center power security a critical concern.

K&L Gates adds AI and cybersecurity expert

K&L Gates has added Sarah Pearce, a cyber, privacy, and data security lawyer, as a partner in its London office. Pearce focuses on regulatory compliance, incident response, and AI governance.