News from the AI & ML world
DeeperML - #aihardware
|
@the-decoder.com
//
Nvidia is making significant advancements in artificial intelligence, showcasing innovations in both hardware and video generation. A new method developed by Nvidia, in collaboration with Stanford University, UCSD, UC Berkeley, and UT Austin, allows for the creation of AI-generated videos up to one minute long. This breakthrough addresses previous limitations in video length, where models like OpenAI's Sora, Meta's MovieGen, and Google's Veo 2 were capped at 20, 16, and 8 seconds respectively.
The key innovation lies in the introduction of Test-Time Training layers (TTT layers), which are integrated into a pre-trained Transformer architecture. These layers replace simple hidden states in conventional Recurrent Neural Networks (RNNs) with small neural networks that continuously learn during the video generation process. This allows the system to maintain consistency across longer sequences, ensuring elements like characters and environments remain stable throughout the video. This new method has even been showcased with an AI-generated "Tom and Jerry" cartoon. Furthermore, Nvidia has unveiled its new Llama-3.1 Nemotron Ultra large language model (LLM), which outperforms DeepSeek R1 despite having less than half the parameters. The Llama-3.1-Nemotron-Ultra-253B is a 253-billion parameter model designed for advanced reasoning, instruction following, and AI assistant workflows. Its architecture includes innovations such as skipped attention layers, fused feedforward networks, and variable FFN compression ratios. The model's code is publicly available on Hugging Face, reflecting Nvidia's commitment to open-source AI development. Recommended read:
References :
Ali Azhar@AIwire
//
References:
AIwire
Nvidia is strategically expanding its AI capabilities with recent acquisitions, signaling a push towards full-stack AI control. The company is reportedly in advanced talks to acquire Lepton AI, a startup specializing in renting out Nvidia-powered servers for AI development. This move, along with the acquisition of synthetic data startup Gretel, demonstrates Nvidia's ambition to move beyond hardware and offer comprehensive AI solutions.
Nvidia's acquisition strategy aims to enhance its cloud-based AI offerings and counter competition from cloud providers developing their own AI chips. The company's interest in Lepton AI and the acquisition of Gretel, known for privacy-safe AI training data, are key steps in its strategy to become a full-stack enabler of AI development. These acquisitions are aimed at integrating into the AI development pipeline and providing more complete solutions for AI development. Recommended read:
References :
Jesus Rodriguez@TheSequence
//
NVIDIA's GTC 2025 showcased the company's advancements in AI hardware and software, solidifying its position as a leader in the AI compute industry. The conference highlighted the synergy between NVIDIA's hardware and software offerings, emphasizing AI's pervasive influence across various sectors. Key announcements included the Blackwell Ultra AI Factory Platform, boasting 72 Blackwell Ultra GPUs and 36 Grace CPUs, designed for demanding AI agent workloads. NVIDIA also previewed future platforms, such as the Rubin Ultra NVL576 slated for late 2027, showcasing their commitment to continuous innovation.
NVIDIA also unveiled the Llama Nemotron family of open-source reasoning models, designed for superior accuracy and speed compared to standard models. These models are already being integrated by major players like Microsoft and SAP into their respective platforms. Furthermore, NVIDIA launched Dynamo, an open-source inference framework aimed at maximizing GPU performance through intelligent scheduling. The event, which attracted an estimated 25,000 attendees, underscored NVIDIA's role as a key enabler of AI advancements, driving innovation from healthcare to autonomous vehicles. Recommended read:
References :
Jesus Rodriguez@TheSequence
//
References:
TheSequence
, BigDATAwire
Nvidia's GTC 2025, held in San Jose, California, concluded this week, showcasing the company's advancements in AI hardware and software. The event, drawing an estimated 25,000 attendees, was described as "The Super Bowl of AI," underscoring Nvidia's dominant position in the high-end GPU market essential for training and running AI models. CEO Jensen Huang, dubbed "AI Jesus," unveiled powerful new hardware like the Blackwell Ultra AI Factory Platform and teased future platforms like Rubin Ultra, demonstrating Nvidia's commitment to meeting the growing compute demands of next-generation AI models.
The conference also highlighted Nvidia's progress on the software front, with the launch of the Llama Nemotron family of open-source reasoning models. These models, designed for accuracy and speed, are already being integrated into platforms like Microsoft's Azure AI Foundry and SAP's Joule copilot. SEEQC and NVIDIA announced they have completed an end-to-end fully digital quantum-classical interface protocol demo between a QPU and GPU. This marks a move towards AI agents capable of solving problems independently. Furthermore, SEEQC and Nvidia reported a breakthrough in quantum computing with a fully digital quantum QPU-GPU interface that leverages Single Flux Quantum (SFQ) technology's ultra-fast clock speeds and on-Quantum Processor digitization to eliminate bandwidth bottlenecks, reduce latency and create an optimal digital link to NVIDIA GPUs. Recommended read:
References :
Chris McKay@Maginative
//
NVIDIA's GTC 2025 event showcased significant advancements in AI infrastructure, highlighting the Blackwell Ultra and Rubin architectures, along with several related technologies and partnerships. Jensen Huang, Nvidia CEO, delivered a keynote address outlining the company’s vision for the AI-powered future, emphasizing improvements in processor performance, network design, and memory capabilities. The Blackwell Ultra GPUs are being integrated into DGX systems to meet the rising demands of AI workloads, especially in inference and reasoning.
NVIDIA is also expanding its offerings beyond chips with the introduction of desktop AI supercomputers for developers. The DGX Station, powered by the GB300 Blackwell Ultra Superchip, aims to bring data center-level AI capabilities to a compact form factor. Nvidia introduced Dynamo, an open-source inference software engineered to maximize token revenue generation for AI factories deploying reasoning AI models. The presentation emphasized a clear roadmap for data center computing, advancements in AI reasoning capabilities, and bold moves into robotics and autonomous vehicles. Recommended read:
References :
@tomshardware.com
//
Nvidia has unveiled its next-generation data center GPU, the Blackwell Ultra, at its GTC event in San Jose. Expanding on the Blackwell architecture, the Blackwell Ultra GPU will be integrated into the DGX GB300 and DGX B300 systems. The DGX GB300 system, designed with a rack-scale, liquid-cooled architecture, is powered by the Grace Blackwell Ultra Superchip, combining 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs. Nvidia officially revealed its Blackwell Ultra B300 data center GPU, which packs up to 288GB of HBM3e memory and offers 1.5X the compute potential of the existing B200 solution.
The Blackwell Ultra GPU promises a 70x speedup in AI inference and reasoning compared to the previous Hopper-based generation. This improvement is achieved through hardware and networking advancements in the DGX GB300 system. Blackwell Ultra is designed to meet the demand for test-time scaling inference with a 1.5X increase in the FP4 compute. Nvidia's CEO, Jensen Huang, suggests that the new Blackwell chips render the previous generation obsolete, emphasizing the significant leap forward in AI infrastructure. Recommended read:
References :
staff@insideAI News
//
Axelera AI has secured a EuroHPC grant of up to €61.6 million to advance the development of its AI chiplet, Titania. The grant is part of the EuroHPC Joint Undertaking's (JU) effort to develop a supercomputing ecosystem in Europe. Axelera AI will support the EuroHPC JU in fostering the design and development of European processors, accelerators, and related technologies for extreme-scale, high-performance, and emerging applications as part of the DARE (Digital Autonomy with RISC-V for Europe) Project.
The Titania chiplet is described as a high-performance, low-power, and scalable AI inference solution. Axelera AI believes that its Digital In-Memory Computing (D-IMC) technology will provide near-linear scalability from the edge to the cloud. The company also said that the chiplet will be enhanced with proprietary RISC-V vector extensions and that its architecture facilitates scaling from the edge to the cloud, streamlining expansion and optimizing performance in ways that traditional cloud-to-edge approaches cannot. This funding follows an oversubscribed $68 million Series B financing round, bringing the total amount raised to more than €200 million in three years. Recommended read:
References :
Rick Champagne@NVIDIA Newsroom
//
References:
NVIDIA Newsroom
, insideAI News
,
NVIDIA is making waves across multiple sectors with its innovative AI technologies. A report indicates that OpenAI and Oracle plan to acquire 64,000 of NVIDIA's Blackwell GB200 chips for the Stargate AI data center project in Texas, which was initially announced in January. This massive purchase underscores the growing demand for NVIDIA's advanced chips in powering large-scale AI infrastructure. The GB200 Superchip, combining a Grace CPU with two enhanced B200 GPUs, is expected to significantly boost the data center's capabilities.
This year, NVIDIA researchers Essex Edwards, Fabrice Rousselle, and Timo Aila were also recognized with Scientific and Technical Awards by the Academy of Motion Picture Arts and Sciences. Their contributions to simulation, denoising, and rendering have revolutionized the film industry, enabling filmmakers to create more realistic and immersive visual experiences. NVIDIA also unveiled "ACE" at CES 2025, promising truly intelligent NPCs for games such as PUBG: Battlegrounds and Naraka: Bladepoint, potentially transforming in-game interactions. Recommended read:
References :
@the-decoder.com
//
References:
bsky.app
, Techstrong.ai
,
OpenAI is finalizing the design of its first custom AI chip, with mass production targeted for 2026. The company is partnering with Taiwan-based TSMC for manufacturing, using their advanced 3-nanometer process. This move mirrors similar efforts by other tech giants to design custom chips for AI workloads, aiming to reduce reliance on NVIDIA and potentially gain leverage in negotiations with suppliers. The chip under development is capable of both training and running AI models, although initial deployment may be limited in scale.
This project is spearheaded by a 40-person team led by Richard Ho, formerly of Google's custom AI chip program, and working closely with Broadcom. OpenAI's endeavor comes as demand for AI chips continues to surge and tech giants like Amazon, Microsoft, and Meta explore their own silicon or third-party options amid rising costs. Microsoft said it will spend $80 billion on AI infrastructure in 2025, Meta has vowed to splurge $65 billion in the next year on building out its AI operations. Recommended read:
References :
@tomshardware.com
//
References:
blogs.nvidia.com
, venturebeat.com
,
The U.S. government is currently investigating whether Chinese AI firm DeepSeek acquired restricted Nvidia GPUs through intermediaries in Singapore. This investigation stems from concerns that DeepSeek may have bypassed U.S. export restrictions on advanced AI hardware. The probe was launched after it was noticed Singapore's share of Nvidia's revenue rose from 9% to 22% in two years, coinciding with increased restrictions on sales to China. DeepSeek's impressive AI models, such as R1, have shown comparable performance to leading models like those from OpenAI and Google, raising questions about the computing power used to train them. It is suspected that restricted Nvidia GPUs, unavailable for export to China, may have been acquired to train these advanced AI models. Nvidia denies any wrongdoing, maintaining that they adhere to all legal requirements.
In other news, Nvidia CEO Jensen Huang recently met with former U.S. President Donald Trump to discuss technology and AI leadership. The meeting took place amid Trump's recent calls for tariffs on foreign-made computer chips, raising concern in the technology industry, with some predicting price increases on consumer electronics. A Nvidia spokesperson stated the discussions centered on strengthening the U.S. position in technology and AI, which comes as DeepSeek’s R1 model has also been integrated into NVIDIA’s NIM platform to help developers utilize it in their own AI agents. DeepSeek's R1 model is also now available on AWS through Amazon Bedrock and Amazon SageMaker AI, enabling broader accessibility and deployment of their powerful AI models for various business applications. Recommended read:
References :
@www.cnbc.com
//
References:
www.techmeme.com
, www.cnbc.com
,
Nvidia's stock plummeted, experiencing a historic single-day loss of nearly $600 billion in market capitalization. This dramatic downturn was primarily fueled by investor concerns surrounding the rise of DeepSeek, a Chinese AI startup. The drop, which saw Nvidia shares fall by 16.86% to close at $118.58, marks the largest one-day market cap loss in US history, more than doubling any previous single day drop for an American company. Anxiety over DeepSeek's advancements has also triggered a sell-off in other US tech stocks pre-market and a slump in Japanese chip stocks, highlighting the widespread impact of the AI competition.
The market's negative reaction stems from DeepSeek's breakthroughs in AI training efficiency, which investors fear could reduce the demand for Nvidia's high-end GPUs. Concerns over DeepSeek are further compounded by a broader narrative of hardware competitors and the ability to translate Large Language Model code to avoid Nvidia's CUDA lock-in. DeepSeek's recent success with its low-cost AI model overtaking ChatGPT on the App store has intensified the worries, sparking discussions about a potential "Sputnik moment" for the US AI industry. Recommended read:
References :
@www.fool.com
//
Quantum computing stocks have dramatically crashed following comments from Nvidia CEO Jensen Huang, who projected that truly useful quantum computers are still 15 to 30 years away. This statement triggered a massive sell-off, wiping out an estimated $8 billion in market value across the sector. Shares of key companies like IonQ, Rigetti Computing, and D-Wave Quantum plummeted, with drops exceeding 30% in a single day. The market reacted negatively to Huang's timeline, undermining previous optimism fueled by breakthroughs like Google's new 105-qubit 'Willow' chip, which was reported to have solved a complex calculation in five minutes, a feat that would take current supercomputers around 10 septillion years to complete.
Despite the setback, some industry leaders are pushing back against Huang's assessment. D-Wave Quantum CEO Alan Baratz dismissed Huang’s comments as “dead wrong,” highlighting that D-Wave's annealing quantum computers are already commercially viable. Baratz emphasized that their technology can solve problems in minutes that would take supercomputers millions of years, challenging Huang's view on current capabilities. He even offered to meet with Huang to discuss what he called “knowledge gaps” in the CEO's understanding of quantum technology. An X user also pointed out that Nvidia is currently hiring quantum engineers, adding further to the industry's resistance to the projected long wait for the technology. Recommended read:
References :
|
BenchmarksBlogsResearch Tools |