Recently, the AI research company Epoch AI released an interactive simulator designed to simulate the computational power required to train large language models. Through this simulator, researchers found that while a 2012 outdated graphics card (like the GTX580) could be used to train GPT-4, the cost would be ten times that of modern hardware today. Epoch AI's research indicates that the number of floating-point operations (FLOP) required to train GPT-4 is between 1e25 and 1e26.

Nvidia has recently engaged in a series of collaborations with local businesses in India to promote its artificial intelligence chips and technologies. The CEO of the American chip design company, Jensen Huang, announced at an AI summit in Mumbai that Nvidia has partnered with one of India's most valuable companies, Reliance Industries, to co-establish the infrastructure for AI applications. Additionally, tech company Tech Mahindra will use Nvidia's chips and software to develop a Hindi AI model called "Indus2.0."

Elon Musk's xAI has announced the launch of an artificial intelligence training system named 'Colossus', powered by 100,000 Nvidia H100 GPUs, which is considered the most powerful AI training system globally. Its performance may even surpass that of the U.S. Department of Energy's Aurora supercomputer. The power of Colossus is largely attributed to Nvidia's H100 GPU, which operates language models at a speed 30 times faster than previous GPUs, especially in transformer-based tasks.

Elon Musk showcased the internal structure of Tesla's Cortex AI supercomputer cluster being built at the Austin headquarters, which is expected to require 130 megawatts of cooling and power upon launch, growing to 500 megawatts by 2026. The cluster contains 70,000 AI servers, primarily used to solve real-world AI issues, particularly for training Tesla's full self-driving system and the AI for the Optimus robot. Cortex will rely entirely on Nvidia's hardware at launch, while Tesla's own hardware is expected.

Nvidia has encountered engineering challenges during the launch of its next-generation AI chip series Blackwell, leading to delays in the product release schedule and affecting market expectations. The main issue arises from the need to redesign an AI accelerator for compatibility with the Hopper H100 chip, as well as a processor combined with a graphics card that cannot be mass-produced due to technical problems. These challenges highlight the complexity of the technology innovation process. As a result, Nvidia's stock price dropped by 6.4%, while competitor AMD's stock price rose by 1.8%. Despite the delays, analysts believe that this...

Meta has trained its latest AI model, Llama 3.1, using 16,384 GPUs, showcasing the astonishing speed of AI technological advancement. However, during this process, an average of one failure occurs every 3 hours, totaling 419 failures, with approximately half related to the H100 GPUs and their HBM3 memory. This data reveals the reliability challenges faced by supercomputing systems in their pursuit of performance breakthroughs. The complexity of the Llama 3.1 training cluster is comparable to a small city's neural network, with frequent failures. The Meta team has implemented strategies such as reducing...

GPU cloud computing company Lambda has launched a 1-Click Cluster Service, allowing users to access Nvidia H100 GPUs and Quantum 2 InfiniBand clusters on demand, especially suitable for companies needing GPU compute power for short-term projects. This innovative service simplifies the coordination of hardware and software, with no long-term contracts required; users simply reserve when needed, with a minimum reservation time of two weeks. Founded in 2012, Lambda recently raised $320 million, valuing the company at $1.5 billion, indicating strong market demand for its services.

Recently, Musk announced on the social platform X (formerly known as Twitter) that his company, xAI, is conducting training on the world's most powerful AI training cluster — the "Memphis Supercluster" located in Memphis, Tennessee. This cluster will be the largest capital investment in the city's history, attracting significant media attention. According to local news reports, this supercluster is equipped with 100,000 liquid-cooled Nvidia H100 graphics processing units (GPUs). These chips, in

On July 9th, Musk announced that his artificial intelligence company xAI is building a supercomputer with 100,000 NVIDIA H100 GPUs, expected to be delivered and begin training by the end of this month. This move marks the termination of xAI's negotiations with Oracle to expand the existing agreement and rent more NVIDIA chips. Musk emphasized that this will become "the most powerful training cluster globally with a significant lead." He said that the core competitive edge of xAI lies in speed, "

Musk spends 4 billion dollars to purchase 100,000 H100 chips, all for training xAI's Grok3! This "Iron Man" of the tech world is not only making waves in the rocket and electric vehicle sectors, but now he's also making a big splash in the AI field. Musk responded twice on Twitter to promote xAI, announcing the release of Grok2 in August and the upcoming Grok3, trained on 100,000 H100 chips by the end of the year. This chip-powered innovative data training aims directly at OpenAI's GPT, aiming t

Elon Musk recently announced that his artificial intelligence startup, xAI, is set to launch its revolutionary large language model, Grok-2, in August, signaling the arrival of more advanced AI capabilities. Although Grok-2 has not yet been fully unveiled, Musk is already excitedly hyping up the upcoming masterpiece, Grok-3. When discussing AI development, Musk particularly emphasized the importance of datasets and the arduous task of cleaning data for large language models (LMMs). He also did n

OpenAI's Sora generates 5 minutes of video per hour on the NVIDIA H100, allowing for 120 minutes of video creation daily to support the TikTok and YouTube creator community, requiring about 720,000 NVIDIA H100 GPUs. Sora will target visual artists and filmmakers after its launch, with plans to incorporate sound and editing flexibility. OpenAI is promoting Sora in Hollywood, allowing creators to generate multiple candidates before uploading videos.

Meta has released two 24K H100 GPU clusters for training Llama-3. Llama-3 utilizes RoCEv2 networking and Tectonic/Hammerspace's NFS/FUSE network storage. It is expected that Llama-3 will go live by the end of April or mid-May, possibly as a multimodal model and will continue to be open-sourced. Meta plans to have 600,000 H100 computing power by the end of 2024.

The delivery time for NVIDIA H100 GPU has been shortened to 3-4 months, and some companies have begun reselling excess H100 80GB processors. Demand for AI chips still exceeds supply, and cloud service providers offer on-demand rental services for H100 GPUs. Prices for NVIDIA H100 and similar processors have not decreased, and companies are placing more emphasis on pricing and procurement rationality. The AI sector may witness a more balanced market landscape.

Nvidia's latest quarterly revenue has seen significant growth, with data center products being the main contributor. In the past quarter, Nvidia sold 500,000 H100 GPUs. Meta and Microsoft are the largest customers of the H100, each purchasing 150,000 GPUs. As companies gradually shift to their own GPUs, sales are expected to decline. Analysts believe Nvidia will reach sales targets for the H100 and A100 again in the fourth quarter.

Microsoft has launched the new ND H100v5 virtual machine series, equipped with 8 NVIDIA H100 chips. The virtual machines feature AI supercomputing GPUs, next-generation CPUs, and low-latency networking. When using the new 8-bit FP8 floating-point data type for matrix multiplication operations, the speed has increased sixfold. ND H100v5 VM helps large language models achieve up to two times the inference speed improvement.