AMD, an American semiconductor company, has launched an advanced GPU for AI applications, potentially posing a challenge to Nvidia’s current market dominance in the AI chip sector. The chip, dubbed MI300X, will begin shipping later this year, with e-commerce titan Amazon as its possible inaugural customer.
The company intends to infiltrate the AI market with alternatives to Nvidia’s offerings. According to AMD’s CEO, Lisa Su, AI represents the company’s most significant and strategic long-term growth prospect.
She discussed the AI accelerator market’s potential growth, citing its possible rise from $30 billion this year to over $150 billion by 2027, given a 50% compound annual growth rate.
Why is AI GPUs Crucial?
Generative AI programs and tools necessitate GPU chips. “GPUs sit at the heart of generative AI,” Su explained to investors in a report by CNBC.
Large language models (LLMs) for generative AI demand substantial memory and processing power due to the burgeoning number of calculations. In a demonstration, AMD ran a 40 billion parameter model called Falcon on the MI300X. To provide context, OpenAI’s GPT-3 model boasts 175 billion parameters.
AMD claims its MI300X chip and CDNA architecture were tailored for large language and high-end AI models. With support for up to 192GB of memory, the GPU can comfortably run larger models, outstripping Nvidia’s H100’s 120GB limit.
Su highlighted the increasing size of models and the need for multiple GPUs to operate them. However, she explained that with the additional memory in AMD chips, fewer GPUs would be necessary.
Is This a Positive Development for AMD?
While AMD has not revealed pricing information, its AI GPU debut could pressure Nvidia to reconsider its AI GPU pricing structure, particularly for the H100 model that can fetch upwards of $30,000. Reduced GPU prices translate to lower costs for running generative AI applications.
Initially, AMD plans to provide an Infinity Architecture, coupling eight of its M1300X accelerators in a single system. It mirrors similar setups by Nvidia and Google that amalgamate eight or more GPUs in a single unit for AI uses.
Amazon Might Employ AMD’s GPU Chips
According to a Reuters report, Amazon Web Services (AWS) is contemplating using AMD’s new AI chips. Still, the leading cloud computing provider has yet to finalize the decision.
“Our teams are jointly working on where precisely AWS and AMD will intersect. It’s still a work in progress,” mentioned Dave Brown, vice president of elastic compute cloud at Amazon. “We’ve seen benefits from the design work they’ve done to integrate with existing systems,” he added.