The artificial intelligence (AI) industry ushered in a crazy night, Google (Google) and AMD have launched new ones one after another, and finally they are about to "blow up" OpenAI and Nvidia.
Titanium Media App news on December 7, Beijing time early this morning, Google CEO Sundar Pichai (Sundar Pichai) announced that Google officially released the most powerful and versatile multimodal artificial intelligence (AI) model to date: Gemini (Chinese called "Gemini").
Specifically, Google's newly released Gemini 1.0 series, mainly English models, includes three different size versions: Ultra (super large cup), Pro (large cup) and Nano (medium cup), which will be fully built-in the latest and most powerful self-developed AI supercomputing chip Cloud TPU v5p.
Google says that the Gemini Pro outperformed GPT-3.5 in six benchmarks, and outperformed GPT-4, the strongest model available, in 30 performance benchmarks. In fact, Gemini Ultra scored as high as 90.0% on the MMLU test, that is, in 57 areas such as mathematics, physics, and law, Gemini Ultra became the first AI model to surpass the level of human experts.
At the same time, in the early morning of December 7, chip giant AMD released the high-profile new Instinct MI300X AI acceleration chip (APU) in San Jose, USA, which is composed of more than ten TSMC 5nm/6nm chiplets, with 153 billion transistors, packaging CPU+GPU, and supporting up to 192GB of HBM3 memory capacity, aiming at the AI computing chip market dominated by NVIDIA.
Pichai said Gemini is a giant leap forward for AI models that will eventually affect almost all Google products. At the same time, humanity has officially ushered in the first step of the new era of Gemini, moving towards a true artificial general intelligence (AGI model).
"The MI300X is our most advanced product to date and the most advanced AI acceleration chip in the industry. AMD CEO Lisa Su said, "AI is the single most important and transformative technology in more than 50 years, and it's not just a cool new technology, it's really changing the future of computing." ”
However, it is a little regrettable that as of now, neither the MI300X nor the Gemini 1.0 series models will be sold and serviced in the Chinese mainland market, and AMD is expected to launch a "China special" compliant product.
Lisa Su, CEO of AMD
Google Gemini surpasses GPT-4 in performance, and the latest self-developed chip TPU v5p model training speed is increased by 280%
Since November 30, 2022, OpenAI has released the AI chatbot ChatGPT, which has taken the world by storm and become one of the fastest-growing consumer applications in the world. Not only is it capable of writing poems and programming, but it can also write screenplays, assist with interviews, publish papers, and even provide a wide range of search services, showing great potential for productivity liberation.
As the company that has invested in AI technology for the longest time and has many core technologies, Google was caught off guard by ChatGPT's excellent performance and technical advantages. The market is concerned that ChatGPT may replace traditional search engines and have an impact on Google's search advertising business.
In response to this challenge, Google launched an internal "code red" alert in February this year, and founders Larry Page and Sergey Brin returned to the company, urgently launched ChatGPT competitor Bard, and invested more than $400 million in ChatGPT competitor Anthropic.
However, despite the fact that Bard has undergone several iterations and upgrades since its release, market feedback has generally been poor, with the perception that its product is not as rushed, performant, and experienced as ChatGPT or Microsoft's GPT-4 version of Bing Search.
Now, Google is finally ready to fight back with the release of Gemini 1.0, the long-rumored "Google killer" and the most powerful general-purpose AI model in history that surpasses GPT-4.
Multimodal Gemini is the largest and most capable Google model to date, surpassing GPT-4 in text, video, voice and other fields, which is a real shame.
According to The Information, Google had planned to announce Gemini in 2024, but now, Google appears to have changed its strategy to seize the spotlight as OpenAI goes through a board reshuffle and internal turmoil. In the latest paper published by Google, Gemini's R&D team has more than 800 people.
Specifically in terms of products, the new Gemini series includes three model versions: Gemini Ultra (extra large cup), Gemini Pro (large cup) and Gemini Nano (medium cup).
Among them, Gemini Ultra is the most powerful model of the three, suitable for highly complex tasks, and is scheduled to be built into Google's chatbot Bard next year, Gemini Pro (Large Cup) is the most versatile and scalable model for various tasks, and is currently in internal testing on Bard, and Gemini Nano (medium cup), the AI model that performs the most efficient tasks mainly in devices, will be used for the first time on the Google Pixel 8 Pro.
In contrast to Bard, Gemini and Bard are both large language models (LLMs) developed by Google AI, but they have different features and benefits. Gemini is better at generating creative text formats such as poems, codes, scripts, musical compositions, emails, letters, etc., while Bard is better at providing information and comprehensive answers to challenging questions. If users need to get information and comprehensive answers, Bard will be a better option.
According to Google's official website, Gemini 1.0 highlights include five main areas: state-of-the-art performance test results, new inference and creative features, a powerful and efficient AI supercomputing system, responsibility and safety, and usability.
First, in terms of performance testing, in 30 of the 32 performance benchmarks, the Gemini Ultra model outperformed the state-of-the-art GPT-4 available. Among them, Gemini Ultra surpassed human experts for the first time with a score of 90.0% in the MMLU (Massive Multitasking Language Understanding) test, which combines 57 subjects such as mathematics, physics, and history, in addition to achieving a high score of 59.4% in the UltraMMMU multimodal test, and in the image benchmark test, Ultra surpassed the powerful GPT-4V without using OCR, showing its multimodal and more complex reasoning ability.
Gemini 1.0 has the ability to understand complex written and visual information with its sophisticated, multimodal reasoning capabilities, and its ability to read, filter, and understand information to extract insights from hundreds of thousands of documents, helping to achieve new breakthroughs at digital speed in fields ranging from science to finance. In addition, after training, Gemini 1.0 can recognize and understand text, images, audio, etc. at the same time, and the coding ability of AlphaCode 2 has been greatly improved, and the encoding performance exceeds 85% of the participating programmers, which is nearly 50% higher than the previous generation of AlphaCode.
In terms of computing power, Gemini 1.0 uses Google-designed TPUs v4 and v5e chips for large-scale AI training. On TPUs, Google said, Gemini runs significantly faster than earlier smaller, weaker models. In addition, today, Google announced Cloud TPU v5p, the most powerful, efficient, and scalable TPU system to date, designed to power the training of cutting-edge AI models that will accelerate the development of Gemini.
It is important to mention that Google's TPU v5p chip uses TSMC's 5nm/7nm process, and compared with the previous generation of TPUv4 inference chips, v5p is more of a training chip developed for generative AI. Each TPU v5p pod (server group) consists of 8,960 chips, which are connected in a 3D topology at 4,800 Gbps per second via the chip-to-chip interconnect (ICI) with the highest TPU bandwidth. Compared to TPU v4, TPU v5p has more than 2x better Flops performance and 3x higher high-bandwidth memory (HBM). At the same time, TPU v5p trains large models 2.8 times faster (280%) faster than the previous generation TPU v4, and with the second generation of SparseCores, TPU v5p trains embedded-dense models more than 1.9 times faster than TPU v41.
In terms of responsibility and safety, Google said it is committed to advancing AI responsibly, especially in the development of Gemini models with advanced multimodal capabilities, adding new protections in line with Google AI principles and strict product safety policies, taking into account potential risks holistically, and testing and risk mitigation at every stage of development. In addition, Google worked with external experts to conduct stress tests to ensure content safety, and built a dedicated security classifier to identify and filter harmful content to ensure that Gemini is more secure and inclusive.
In terms of usability, Google announced that starting today, Bard will use a fine-tuned version of Gemini Pro for more advanced reasoning, planning, and understanding, and will be available in English in more than 170 countries and regions (excluding Chinese mainland), as well as plans to expand to different models and new languages and locations in the future; in addition, developers and enterprise customers can access Gemini through Google AI Studio or Google Cloud Vertex AI from December 13 Pro's API, while the Gemini Nano is built into the Pixel 8 Pro.
Google says Gemini will be applied to more of Google's products and services, such as search, ads, Chrome and Duet AI, in the coming months. According to testing, with Gemini built into Google Search, it is able to provide users with a faster search generation experience (SGE) and a 40% reduction in user latency in English searches in the United States, while also improving quality.
谷歌CEO Sundar Pichai（来源：CNBC）
Pichai said the launch of Gemini was an important moment for Bard and the beginning of the Gemini era. If the test results are accurate, the new model may have made Bard as good as ChatGPT. "This is obviously already quite an impressive feat. ”
However, as of now, Google's search and advertising business has not been affected by large models such as GPT. According to Statcounter data, as of October 2023, Bing's market share in the U.S. was just under 7%, down from 7.4% in the same period last year, while Google continued to dominate the search engine market with an 88.13% share during the same period. So far in 2023, the market share of Bing search equipped with GPT technology has remained below 7%. At the same time, the financial report shows that in fiscal year 2022, among the total revenue of Google's parent company Alphabet, Google search and other revenues will be $162.5 billion, a year-on-year increase of 9%. In the first three quarters of fiscal 2023, Alphabet's total revenue reached $221.084 billion, exceeding the $206.8 billion in the same period last year.
"We've made tremendous progress with Gemini so far. We're working to further expand the various features of its future releases, including making progress in planning and memorization, as well as processing more information and providing better responsiveness by adding contextual windows. Google said.
According to The Verge, Google plans to launch a preview version of "Bard Advanced" powered by Gemini Ultra next year, Google's most powerful Chat product in history, which is expected to benchmark ChatGPT-4 or 5.
But Pichai also admitted that although AI will be more significant to human beings than the birth of fire and electricity, Gemini 1.0 may not change the world, and at best, Gemini may help Google catch up with OpenAI in the generative AI arms race.
"This is an important milestone in the evolution of AI and marks the beginning of a new era for Google, where we will continue to innovate rapidly and responsibly improve the capabilities of our models. Google said.
The most powerful AI chip, MI300X, has 60% higher performance than Nvidia H100, but it is not sold in China
AMD has finally decided to fight back against Nvidia.
(Image source: AMD)
In the early morning of December 7, Beijing time, at the AMD Advancing AI event held in San Jose, USA, AMD CEO Lisa Su (Lisa Su) announced the launch of the Instinct MI300X AI acceleration chip (APU) and announced the mass production of the MI300A chip.
Both products are aimed at this NVIDIA-dominated market.
Among them, the MI300X memory is 2.4 times that of Nvidia's H100 products, and the memory bandwidth is 1.6 times that of H100, which further improves performance and is expected to challenge NVIDIA's position in the hot AI acceleration chip market.
Specifically, AMD said that the new MI300X chip can improve the performance of NVIDIA's H100 by up to 60%. The MI300X performance is up to 20% better in a one-to-one comparison with the H100 (Llama 2 70 billion parameter version), a 20% improvement in performance with the H100 (FlashAttention 2 version), a 40% improvement in performance with the H100 (Llama 2 70B version) in an 8-on-8 server comparison, and a 40% improvement in performance with the H100 (Bloom 176B), the MI300X delivers a 60% performance improvement.
At the same time, in AI large model training, compared with H100, MI300X improves the BF16 performance benchmark by 3.4 times, INT8 accuracy performance by 6.8 times, and FP8 and FP16 TFLOPS by 1.3 times, thus further improving the training performance.
Su said that the new MI300X chip is comparable to the H100 in terms of its ability to train AI software, and is much better than the H100 in terms of inference, that is, the process of running the software after the software is put into actual use.
For AI, AMD has three key strengths: complete IP and broad compute engine portfolio to support the most demanding workloads from cloud and edge to endpoint, the company is expanding its open source software capabilities to lower the barrier to entry and use of AI's full potential, and AMD is deepening its AI partner ecosystem to enable cloud service providers (CSPs), OEMs and independent software developers (ISVs) to enjoy its pioneering innovations.
At present, AMD, NVIDIA, and Intel are fully promoting the AI boom. Among them, Nvidia has announced the 2024 Hopper H200 GPU and Blackwell B100 GPU product information, and Intel will launch the Guadi 3 and Falcon Shores GPUs in 2024, and the three companies are expected to continue to compete in the coming years.
AMD said that Dell, HP, Lenovo, Meta, Microsoft, Oracle, Supermicro and other companies will use the MI300 series products.
Among them, Microsoft has partnered with AMD to launch the new Azure ND MI300x v5 virtual machine (VM) series, which is optimized for AI workloads and powered by the MI300X, and Meta will adopt AMD's new MI300X chip products in the data center. Oracle said it will use AMD's new chips in cloud services, which are expected to be released in 2024.
However, AMD is currently facing three main problems: first, AMD production capacity is limited, Nvidia and Intel and other manufacturers have bought production capacity in advance, and AMD has launched MI300A, MI300X and other series of products, or there will be a shortage of supply; second, Nvidia will release B100 in March next year, and H200 will also be sold to the market, so AMD has exclusive "strongest AI chips" Third, in terms of software system, CUDA has become the best choice for large model training and inference, although AMD has made significant progress in the field of ROCm software, but compared with NVIDIA's CUDA ecological ROCm is still very weak.
AMD responded: The Chinese market is important to AMD, but today it did not announce a special product specifically for the Chinese market.
Previously, the market expected AMD's MI300 series to ship about 30~400,000 units in 2024, and the largest customers were Microsoft and Google.
"AMD doesn't think it needs to beat Nvidia to do well in the market," Su said publicly this morning, "We believe we can get a slice of the more than $400 billion AI chip market by 2027." ”
After the MI300X news was released, the stock prices of AMD and Nvidia did not rise but fell.
As of the close of the U.S. stock market on December 6, Nvidia shares fell 2.28%, and AMD shares also fell 1.32% to close at $116.82 per share. Year-to-date, AMD shares are up 82.47%, outperforming Intel but nowhere near Nvidia's shares, which are up 220%.
Has the time really come for AI to surpass humans?
Lisa Su said that AI has not only developed rapidly in a short period of time, but has also shown explosive growth, which is comparable to the invention status of the Internet. But the difference is that AI adoption will be faster. However, we are still in the very early stages of the new era of AI.
The crazy results of AMD and Google this night tell us that the era of AI has arrived, and humanity must pay attention. Just over a year later, the market has undergone a huge shift and turnaround, and anything can happen.
As 2023 enters the last month, the "battle for the throne" over the big model is even more intense.
A year ago, Lisa Su believed that the AI accelerator market in 2023 would be $30 billion. The global addressable market for data center AI accelerators (TAMs) will reach $150 billion by 2027, implying a compound annual growth rate (CAGR) of around 50% during the period.
But now, Su Lifeng directly gave a higher expectation, in 2023, the potential market size of AI accelerators will reach $45 billion, the CAGR in the next few years will be as high as 70%, and by 2027, the entire market will increase to the scale of $400 billion.
When Google's products surpass GPT-4 and AMD's products surpass Nvidia's H100, the AI achievements invented by humans are changing the world, and mankind is ushering in the eve of great changes.
In 2014, British physicist Stephen Hawking issued a stern warning, "The full development of artificial intelligence could mean the end of humanity." ”
Now, the era of AI is finally here, and machines may soon replace humans. From healthcare to the automotive industry, AI technology is being applied across a wide range of industries to profoundly change everyone's lives. In the future, we will see more AI innovations, and we can also foresee a surge in R&D and competition with more cutting-edge technologies.
"Now, [AI] isn't smarter than we are. But I think they [AI] may soon surpass humans," said Geoffrey Hinton, an AI pioneer and Turing Award winner.
(This article was first published on the Titanium Media App, author: Lin Zhijia)