AI Chip Cousins Square Off (NVIDIA vs AMD)

NVIDIA and AMD Square Off in Fight to Take Control of AI

From Data Centers to Laptops, CEOs Huang (NVIDIA) and Su (AMD) Unveil Cutting-Edge AI Technologies

At the Computex summit in Taipei, Nvidia's CEO Jensen Huang and AMD's CEO Lisa Su unveiled their latest AI chips, emphasizing their strategic visions to dominate the AI market. Nvidia highlighted its new "Rubin" chip, slated for 2026, which will succeed the Blackwell series and enhance AI capabilities with improved efficiency. AMD, on the other hand, showcased neural processors designed for AI tasks on laptops, aiming to enhance battery life and performance. Despite their different approaches, both companies are accelerating their product development timelines to secure their positions in the rapidly evolving AI landscape.

Sponsor

Vultr is empowering the next generation of generative AI startups with access to the latest NVIDIA GPUs.

Try it yourself when you visit getvultr.com/forwardfutureai and use promo code "BERMAN300" for $300 off your first 30 days.

• OpenAI Is Rebooting Its Robotics Team - OpenAI is reviving its robotics division, as per Forbes reporting. The AI research firm had previously closed this wing in 2020, citing insufficient training data as the reason. But now, amidst a surge of investments in AI robotics, OpenAI is again recruiting for its robotics team. The company aims to support rather than compete with robotics manufacturers, focusing on developing technological integrations. OpenAI's interest in robotics spans back to its early days, with significant past endeavors like a successful neural network-trained robotic hand-solving a Rubik’s Cube. While OpenAI hasn't specified whether it will re-enter hardware development, it remains committed to advancing AI through collaboration and leveraging its expertise in areas like reinforcement learning and AI safety.

  

• Microsoft CEO Satya Nadella is reportedly worried about an OpenAI deal with Apple - OpenAI's CEO, Sam Altman, is reported to have finalized an arrangement with Apple to integrate OpenAI services into Apple’s products. This move concerns Microsoft's CEO, Satya Nadella, due to its potential effects on Microsoft's ambitions. While OpenAI, Apple, and Microsoft did not comment on the claims, this deal suggests a shift in AI service partnerships, as Apple also considered Google. Amidst this business development, Altman faces internal scrutiny. Ex-board member Helen Toner accused him of deceit and lack of transparency, further highlighted by Jan Lieke's recent resignation from OpenAI, criticizing the company's prioritization of product development over safety and ethical processes. These challenges underscore tension and strategic realignment within OpenAI, influenced by stakeholder expectations and pressures.

• US-returned Chinese physicist and team achieve world first in quantum computing - Chinese scientists, led by quantum physicist Duan Luming, have significantly advanced quantum computing by building the world's largest quantum simulation machine using trapped ions, surpassing previous efforts with a two-dimensional crystal of 512 ions. This architecture is key for scalable, universal quantum computers. Duan, formerly of the University of Michigan and now at Tsinghua University, has made this leap after his extensive experience in the US. The team's 300-ion qubit simulation demonstrated computational possibilities far beyond classical computers, indicating a major step forward in the field.

• Nvidia Unveils Next Generation of AI Chips in Bid to Entrench Market Lead - At the Computex event in Taipei, Nvidia's CEO Jensen Huang revealed the new "Rubin" AI processors, succeeding the "Blackwell" chips announced in March. This move aims to strengthen Nvidia's dominance in the AI chip market and address energy efficiency concerns in data centers. Nvidia also introduced the Vera Arm-based CPU and announced partnerships with PC makers Asus and MSI to enhance AI capabilities in laptops, further solidifying its position in the AI-driven computing age.

  

• Ansel Adams Estate Takes Adobe to Task Over A.I.-Generated Images - The Ansel Adams estate publicly criticized Adobe for allowing AI-generated images in the style of Ansel Adams to be sold on Adobe Stock, despite company guidelines against the use of artists' names. Although Adobe removed the contentious listing after the estate's complaints were aired on Threads, the incident sparked debate about AI art's copyrightability and Adobe's responsibilities to creators. The estate insists on prohibiting the unauthorized use of Adams's name for selling any products, emphasizing the distinct issues between using his name in prompts and training models on his work. Adobe has not commented directly on the issue, but stated action was taken for the violation of Stock terms.

• The CEO of Zoom wants AI clones in meetings - Zoom CEO Eric Yuan revealed the company’s plans to introduce AI clones and digital twins to enhance videoconferencing capabilities. These innovations aim to alleviate “Zoom fatigue” by allowing virtual representations of users to attend meetings and handle tasks like answering emails. Zoom is expanding its product lineup to become an integrated workday operating system, reflecting Yuan's vision of making Zoom meetings more effective than in-person interactions. This approach aims to maintain Zoom's market position amid increasing competition and evolving user expectations​.

  

• AMD Announces Future AI Chips, Will Speed Rollout of New Models - AMD CEO Lisa Su announced that the company will accelerate the release of new AI processors to challenge Nvidia's dominance in the AI market. The MI325X, set for release in Q4 2024, will feature enhanced memory and data throughput. It will be followed by the MI350 in 2025 and the MI400 in 2026, aligning with Nvidia's annual update cycle. Additionally, AMD introduced its third-generation Ryzen AI processors, Strix Point, tailored for laptops, and new gaming processors, including the Ryzen 9 9950X. AMD aims to increase its AI accelerator sales to $4 billion this year, although this remains small compared to Nvidia's projected $100 billion from its data center unit.

• Selling Data for AI May Be Publishers’ Salvation - Internet publishers are increasingly licensing their content to developers of large language models (LLMs), which could become a primary revenue source, surpassing traditional advertising. Companies like Reddit, Shutterstock, and Freepik have struck deals to provide data for AI training, with Reddit expecting significant revenue growth from these arrangements. As tech giants like OpenAI and Google seek high-quality, real-time data, publishers are adapting their strategies to monetize their content. This shift might transform websites, potentially reducing reliance on ads. However, this new revenue model raises concerns about data privacy, user consent, and regulatory oversight, highlighting the evolving dynamics between content creators, publishers, and AI developers.

• Windows feature that screenshots everything labeled a security “disaster” - Microsoft's upcoming Recall feature on Copilot Plus PCs has been under scrutiny for its cybersecurity implications. Recall uses AI to take constant screenshots for easy retrieval, with all data intended to remain on-device. However, cybersecurity expert Kevin Beaumont found that screenshots are stored in plain text within an easily accessible local database, raising the risk of data theft through malware. Despite encryption measures, only data at rest is truly protected, meaning active sessions on stolen laptops would still be vulnerable. Privacy concerns have been raised, considering the feature is enabled by default and lacks an intuitive opt-out during setup. Microsoft has yet to fully address these issues, despite emphasizing security as a company priority.

  

Awesome Research Papers

• FineWeb: Decanting the Web for the Finest Text Data at Scale - Hugging Face introduces FineWeb, a tool designed to efficiently curate high-quality text data from the web. This innovation aims to enhance machine learning models by providing them with cleaner, more relevant data sets. FineWeb's capabilities include filtering and extracting valuable information from vast amounts of online content, making it a powerful resource for improving AI model training and performance.

• 4Diffusion: Multi-view Video Diffusion Model for 4D Generation - The paper introduces 4Diffusion, a novel pipeline for generating spatial-temporally consistent 4D content from single-view videos. It incorporates a learnable motion module into a frozen 3D-aware diffusion model to capture correlations across multiple views over time. A curated dataset is used to train the model, which maintains the 3D model's generalizability and spatial consistency. Additionally, the proposed 4D-aware Score Distillation Sampling loss and anchor loss enhance 4D representation and detail fidelity, using dynamic NeRF to integrate various prior knowledge and address temporal inconsistencies. Comparative experiments show that 4Diffusion outperforms previous methods in generating coherent 4D content.

• Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality - The abstract discusses the relationship between Transformers and state-space models (SSMs), particularly highlighting a model named Mamba which rivals or surpasses Transformers in certain scenarios. The research introduces a theoretical framework, termed SSD, that links SSMs with attention variants via structured semiseparable matrices. Leveraging this framework, the authors created Mamba-2, an enhanced SSM architecture. Mamba-2 is notable for its performance speed, being 2-8 times faster than its predecessor, while maintaining competitive language modeling capabilities akin to Transformers.

State Space Duality (Mamba-2) Part I - The Model - Goomba Lab introduces Mamba-2 with its new Structured State Space Duality (SSD) model. SSD aims to bridge state space models (SSMs) and attention mechanisms while enhancing computational efficiency through matrix multiplications. It features a scalar-times-identity structure for improved training on GPUs and TPUs. The model shows potential performance improvements over Mamba-1, with significant architectural changes to optimize training and expressivity. Future blog posts will delve deeper into the theoretical and practical aspects of the SSD framework.