AMD Launches New Instict MI300X AI Chip To Compete With Nvidia’s GPUs

Chipmaker Advanced Micro Devices Inc. takes on Nvidia Corp. with the launch of a new artificial intelligence accelerator chip announced at today’s Data Center & AI Technology Premier Event.

In her keynote, AMD Chairman and CEO Lisa Su announced the company’s new Instinct MI300X accelerator specifically targeted at generative AI workloads. Generative AI is the technology behind OpenAI’s famous chatbot ChatGPT, which responds to questions and prompts in a human-like conversational way.

Su previously described the MI300A as “the world’s first data center-integrated CPU + GPU” in his keynote at the 2023 Consumer Electronics Show in January, arguing that both the central processing unit and the graphics processing unit could be combined into a single processor. said to be integrated. Today, customers will be able to use her MI300X to run generative AI models with up to 80 billion parameters, she said. Customers will be able to get samples of the chips in the third quarter, and production is expected to ramp up by the end of the year.

AMD, like Nvidia, said it sees a big opportunity in the AI ​​space swept away by the ChatGPT hype. This is because the technology requires a large amount of computing power in the data centers that host it. Su reiterated that today, saying AI is the company’s “most strategic long-term growth opportunity.”

“We are thinking about AI accelerators in the data center [market] It will grow from around $30 billion this year to more than $150 billion in 2027 at a compound annual growth rate of more than 50 percent,” Su said.

Along with MI300X, Su introduced AMD’s Infinity Architecture Platform, a self-contained platform for running generative AI inference and training workloads. He said it is powered by eight MI300X chips and rivals Nvidia’s DGX supercomputer platform for AI applications.

Analysts have had high hopes for AMD since it acquired chip maker Xilinx in February 2022 for $49 billion. During the keynote, AMD President Victor Peng, who previously led Xilinx as CEO, introduced the company’s ROCm software ecosystem for data centers. accelerator. ROCm aims to support the MI300X accelerator and “unify the open AI software ecosystem,” Peng said.

AMD is once again trying to counter Nvidia’s AI strategy, which relies not only on the chip’s hardware, but also on the software ecosystem that allows companies to get the most out of their AI chips.

AMD’s hardware and software are undoubtedly attractive, but the company’s stock fell more than 3% today in the absence of a major customer announcement for a smaller version, dubbed the MI300X, or MI300A, according to analysts. . Traditionally, AMD has always featured big companies early adopters of their new hardware, but today they weren’t able to do so. The company also didn’t provide any pricing details for the MI300X.

“It may have disappointed the public that[big customers]didn’t say they would use the MI300A or X,” Tirias Research analyst Kevin Crewell told Reuters. . “They wanted AMD to say they replaced Nvidia in some designs.”

Nvidia now completely dominates the AI ​​computing industry, with a market share of 80% to 95%, analysts say. The excitement over AI has sent Nvidia’s stock soaring, making it the first chipmaker to hit a $1 trillion market cap late last month. There are few significant competitors in this field. Both Intel Corp. and startups such as SambaNova Systems Inc. and Cerebras Systems Inc. have competing hardware offerings but have failed to gain significant traction. Nvidia’s biggest competitors are probably Alphabet Inc.’s Google Cloud and Amazon Web Services Inc., both of which develop in-house AI chips and make them available for rent through their respective cloud infrastructure platforms.

Not everyone was so critical of AMD, though. Moor Insights & Strategy’s Patrick Moorhead told SiliconANGLE that it’s important to look at the long-term outlook, saying the hyperscale cloud AI accelerator market is likely to be worth $125 billion over the next five years. It pointed out. He said cloud hyperscalers rely too much on Nvidia and are uncomfortable with the lack of a broader supply chain, so that’s where AMD can make a difference.

“Long term, cloud players can either make their own chips, continue to source exclusively from NVIDIA, or work with a second player like AMD,” Moorhead said. said. “If AMD passes the 3rd quarter sampling, I believe it will be able to start its AI business step by step between the 4th quarter of 2023 and the 1st quarter of 2024. In addition, I believe that we will be able to launch AI businesses like PyTorch. AI framework providers and underlying model providers like Hugging Face want: They also have more options to power their innovations.”

Moorhead also pointed out that despite Nvidia’s dominance, most AI inference workloads still run on CPUs rather than GPUs, and that’s unlikely to change anytime soon. “AI spans from the smallest his IoT devices to smartphones, PCs and the largest hyperscale data centers,” he said. “That AI will be distributed across the CPU, GPU, NPU up and down the value chain.

In another announcement today, Su also introduced AMD’s 4th generation Epyc datacenter CPUs (codenamed Bergamo), which will compete with CPUs from Intel and Nvidia.

AMD is much more competitive in the traditional data center space and the Bergamo chip seems to have very good specs. Su said his CPU with 128 cores per socket is better in terms of energy efficiency despite using fewer servers. She also said that she has 1.8x better performance per watt than Intel’s 4th Generation Xeon Platinum CPUs.

AMD has at least a major customer for the Bergamo chip, and Alexis Bjorlin, vice president of computing infrastructure at Meta Platforms Inc., said the company is already using the Bergamo chip as an early adopter.

Image: AMD