Edit: Amao
Design: Mu You
In December 2022, the AI startup OpenAI launched the chatbot ChatGPT. As one of the practical applications of generative AI in the field of text, ChatGPT was only a hundred days ago, but it has become popular all over the world.
Don’t you see, college students use it to write papers, business people use it to formulate proposals, poets use it to create poems, programmers use it to write code... Netizens lamented: "Only what you can’t think of, what can’t be done without ChatGPT. "
Let the real data speak for itself—less than two months after its launch, the global active users of ChatGPT have exceeded 100 million. Such an astonishing growth rate really shattered the glasses all over the place.
For a time, a large number of enterprises have invested in the field of generative AI, with the posture of "letting the past change into the sea, and one pill will make a thousand springs". Some people ask, creative creation can no longer stop artificial intelligence, will thousands of industries be completely subverted and restructured in the future?
The wish is beautiful, but the reality is very skinny. It is a pity that in the final analysis it is still the same sentence: Without powerful and economical computing power, nothing can be discussed.
As Adam Selipsky, CEO of Amazon Cloud Technology, puts it: “Generative AI has the potential to transform entire industries, but the cost and expertise it requires make it inaccessible to all but a few companies.” technology."
A great breakthrough from the beginning
Our world is in an era of explosive computing power.
Taking the development of large AI models such as ChatGPT as an example, the demand for computing power will roughly double every two months. Such a growth requirement is difficult to meet even with the doubled Moore's Law, not to mention the huge cost requirement.
We see that giants like Google are doing similar technology. However, if ChatGPT is deployed to Google Search, the cost will exceed $100 billion, even if only considering the server and network upfront costs.
At this time, we have not considered the later operation and other costs. According to estimates, a single reply from ChatGPT costs at least 1 cent. Before it became a "popular model", the company needed at least $100,000 in operating costs per day to meet the needs of more than 1 million users at the time. However, when the number of users reaches hundreds of millions, the annual cost will exceed US$5 billion.
Obviously, such a high cost is basically unacceptable for any enterprise, even if there is a mine at home, it cannot withstand such a toss. In this case, if you want to increase the computing power by hundreds or thousands of times, you must need a new computing power platform.
On February 28, the market research organization Counterpoint released the latest server CPU survey report. One of the highlights of the report is that, in addition to the traditional two chip manufacturers, Amazon Cloud Technology ranks third, with a market share of 3.16%, nearly double that of 2021.
Although it seems that there is still a relatively large gap with the leaders, all great breakthroughs in the field of science and technology start from the smallest. The example of "a single spark can start a prairie fire" has actually been repeatedly verified in the computing field.
It should be pointed out that since the first generation of Amazon Graviton, Amazon cloud technology has not been sold to the outside world, but has been completely used to provide cloud services to the outside world. That is to say, through Graviton, in addition to traditional computing power, Amazon Cloud Technology has brought new and more inclusive computing power to customers.
At the Amazon Cloud Technology 2022 re:Invent Global Conference held at the end of last year, Graviton, an Arm-based CPU chip developed by Amazon Cloud Technology, released an enhanced third-generation version, Graviton3E, and used it for more computing instances.
Graviton is not only used in the famous EC2 (Amazon Elastic Compute Cloud) , but more hosting services of Amazon cloud technology, such as the latest containers, are based on Graviton. In fact, when using these services, many customers have already clearly felt the cost-effective improvement brought by Graviton.
The key to the implementation of artificial intelligence
There is no doubt that machine learning is profoundly affecting and driving breakthroughs that affect every aspect of our work and life. From traditional enterprises to innovative enterprises, every company is using machine learning technology to solve propositions related to survival and development.
"Whether it's intelligent voice, autonomous driving, or the recent hot AI painting, all AI developments have made machine learning models more and more complex." In a recent exchange, a technical expert at Amazon Cloud Technology said .
True. In the past few years, the scale of the model has continued to expand, and the number of parameters has increased from hundreds of millions to hundreds of billions. The high cost of training and deploying these increasingly complex machine learning models is driving many enterprises, especially small innovative companies shut out.
According to reports, since 2017, the engineer team of Amazon Cloud Technology has noticed such a trend. In their view, if the customer's needs are not met as soon as possible, the high cost of machine learning will soon become unbearable for the customer.
To this end, Amazon Cloud Technology designed and launched Amazon Inferentia in 2019 to provide high performance for machine learning applications, allowing customers to enjoy AI dividends while also being affordable.
In terms of chip design and construction, Amazon Cloud Technology has a top team in the industry, and has developed many excellent products during more than ten years of development, such as the Graviton series, Inferentia, and Amazon Nitro systems we mentioned earlier.
As a server chip dedicated to machine learning inference and drive, Inferentia provides better price/performance ratio, higher throughput and lower latency than similar GPU-based servers.
Also at the 2022 re:Invent Global Conference, Amazon Cloud Technology launched the Inferentia2 chip and Inf2 instance. Compared with the previous generation, the performance per watt of the new Inf2 instance is increased by 45%, the throughput is increased by 4 times, and the delay is only 1/10. It can support ultra-large complex deep learning models with up to 175 billion parameters. GPT type model? Just let the horse come!
Although the chip performance of machine learning has made great progress, it is still difficult to keep up with the increase in training complexity. A feasible solution is to use distributed multiprocessors to perform collaborative computing and collaborative training through the network.
To this end, Amazon Cloud Technology has also specially built an Amazon Trainum chip for machine learning. After being equipped with 16 Trainums, 512GB of accelerator memory and 800GBps of network bandwidth, the Trn1 instance fully demonstrates its power:
Training costs are reduced by 50% compared to similar GPU-based instances. Taking a two-week training of a large model with trillions of parameters as an example, the GPU server P3dn needs 600 instances, the latest generation GPU instance P4d needs 128 instances, but Trn1 only needs 96 instances, so the economy can also be seen One spot.
At the conference at the end of last year, Amazon Cloud Technology also launched Trn1n, a network-optimized instance based on Trn1, which increases the network bandwidth from 800GBps to 1600GBps, and can build more than 10,000 Trainium chips in a super-large-scale cluster...
By delivering powerful performance at an affordable cost, Amazon Cloud Technology opens up new avenues for customers to innovate.
Competing in the digital economy is computing power
The popularity of artificial intelligence is not just ignited by the recent ChatGPT. In fact, last year game designer Jason Allen won the first prize in the art competition for his paintings created through the Midjourney platform, which is one of the vivid examples in our memory.
In the work, titled "Space Opera House," the sun shines through a huge round window into the entire hall, and in a baroque palace, several women in luxurious classical costumes look out into space.
However, "Space Opera House" is the work of artificial intelligence. Although there is no shortage of voices in the industry that "art and creative work should not be affected by machines", it does not affect the arrival of a wave of AIGC in the slightest.
AIGC (AI Generated Content, Artificial Intelligence Automatically Generated Content) is a new content creation method after professionally produced content (PGC, Professional-generated Content) and user-generated content (UGC, User-generated Content) .
In terms of creativity, expressiveness, iteration, dissemination, and personalization, AIGC can give full play to its technological advantages to create new forms of digital content generation and interaction, the existing AI painting, AI writing, and the aforementioned ChatGPT, etc. All belong to the specific manifestations of AIGC.
Of course, AIGC is not only used for chatting and drawing. After a lot of exploration and experimentation, it will inevitably turn to more valuable industrial application fields, thus exerting a huge influence on economic and social development.
To some extent, AIGC is not just technological innovation or application innovation, but a brand new business model innovation. In the 2023 annual report written by the team led by "Sister Wood" Cathie Wood, 12 investment themes with promising prospects were listed, among which AIGC was listed.
Driven by AIGC, enterprises' demand for high computing power has seen a new round of exponential growth—everyone understands that strong and economical computing power support has become a necessary prerequisite for the in-depth development of AIGC, which can provide a new round of opportunities for business innovation.
"It is better to retreat and form a net if you are envious of fish in Yuanyuan." In fact, in the face of the huge business opportunities brought by AIGC, enterprises do not need to "build a net" by themselves. Focusing on various needs, Amazon cloud technology has already woven a powerful computing power network, so that the dividends of the digital age can benefit thousands of enterprises.
In this way, through continuous innovation in chips and services, Amazon Cloud Technology helps customers to better understand and explore AIGC's practices in various fields, and realize AIGC's key trend insight and rapid implementation.
With the continuous and rapid development of my country's digital economy, new business forms, new models, and new applications like AIGC are still emerging rapidly, and they without exception put forward higher requirements for computing power. By providing powerful, economical and green general-purpose and intelligent computing power, Amazon Cloud Technology is committed to the business success of more customers, and fully supports and promotes the high-quality development of China's digital economy.
Click "Read the original text" to download the "Next Generation Cloud Infrastructure White Paper"