Challenge NVIDIA! AMD launches AI chip that can run larger models and write poetry

The number two player in the AI ​​computing power market, chip manufacturer AMD, has launched a new artificial intelligence GPU MI300 series of chips to compete with Nvidia in the artificial intelligence computing power market.

In the early morning of June 14th, Beijing time, AMD held the "AMD Data Center and Artificial Intelligence Technology Premiere" as scheduled, and launched the AI ​​processor MI300 series at the meeting. MI300X, specifically optimized for large language models, will ship to a select group of customers later this year.

AMD CEO Su Zifeng first introduced MI300A, which is the world’s first accelerated processor (APU) accelerator for AI and high-performance computing (HPC). There are 146 billion transistors spread across 13 chiplets. Compared with the previous generation MI250, the performance of MI300 is eight times higher and the efficiency is five times higher.

Subsequently, Su Zifeng announced the most watched product of this conference-MI300X, which is a version optimized for large language models.

"I like this chip," Su Zifeng said. The MI300X chip and its CDNA architecture are designed for large language models and other cutting-edge artificial intelligence models.

Large language models for generative AI applications use large amounts of memory as they run more and more calculations. AMD demonstrated the MI300X running a 40 billion parameter Falcon model. OpenAI’s GPT-3 model has 175 billion parameters.

She also used Hugging Face’s large model based on MI300X to write a poem about San Francisco, where the event was held.

Su Zifeng said that the HBM (high bandwidth memory) density provided by MI300X is 2.4 times that of NVIDIA H100, and the HBM bandwidth is 1.6 times that of competing products. This means AMD can run larger models than the Nvidia H100.

"Model sizes are getting larger and larger, and you actually need multiple GPUs to run the latest large language models," Su Zifeng pointed out. With the increase in memory on AMD chips, developers will not need as many GPUs, Can save costs for users.

AMD also said that it will launch an Infinity Architecture that integrates 8 M1300X accelerators in one system. Similar systems are being developed by Nvidia and Google, which combine eight or more GPUs into a box for artificial intelligence applications.

A good chip is not only the product itself, but also needs a good ecosystem. One reason AI developers have historically favored Nvidia chips is that it has a well-developed software package called CUDA that gives them access to the chip's core hardware functionality.

In order to benchmark NVIDIA's CUDA, AMD launched its own chip software "ROCm" to create its own software ecosystem.

Su Zifeng also told investors and analysts that artificial intelligence is the company’s “largest and most strategic long-term growth opportunity.” “We believe that the data center artificial intelligence accelerator (market) will grow by more than 50% The compound annual growth rate will grow from about US$30 billion this year to more than US$150 billion in 2027.”

Large language models such as ChatGPT require the highest performance GPUs for computing. Nowadays, NVIDIA has an absolute advantage in this market, owning 80% of the market, while AMD is regarded as a strong challenger.

Although AMD did not disclose the price, this move may put price pressure on Nvidia GPUs, such as the latter's H100, which can reach more than 30,000 US dollars. Falling GPU prices may help reduce the high cost of generating AI applications.

However, many of AMD's products at this conference were in a "leading" position in terms of performance. However, the capital market did not push up AMD's stock price. Instead, it closed down 3.61%, while its peer Nvidia closed up 3.90%, its market value for the first time. Closed above the $1 trillion mark.

[Source: The Paper]

The above is the detailed content of Challenge NVIDIA! AMD launches AI chip that can run larger models and write poetry. For more information, please follow other related articles on the PHP Chinese website!