[Tech Talk] Even if You Missed Nvidia... 'Mabel' Still Remains

AI Processing's Another Challenge 'Data Communication' Solved
Marvell's DPU Emerges as Data Center's 'Third Chip'
Competitors All Acquired by Nvidia and AMD

Wall Street in the United States has identified 'Marvell Technology' as another semiconductor beneficiary of the artificial intelligence (AI) boom. While Nvidia, which produces graphics processing units (GPUs)?the core semiconductors for AI acceleration?is currently receiving the most attention, investors who missed out on Nvidia can consider Marvell as an alternative.

Like Nvidia, Marvell is a semiconductor design company. However, unlike products such as central processing units (CPUs) or GPUs, Marvell makes chips that facilitate data transmission among the numerous semiconductors installed in data centers. This distinction makes Marvell another 'AI beneficiary stock.'

Not CPU or GPU... Marvell Rides the AI Boom with 'DPU'

Marvell's Data Processing Unit (DPU) is a key semiconductor for data center operation and artificial intelligence (AI) processing.
[Image source=Marvell]

Marvell, listed on Nasdaq, saw its stock price rise more than 61% last month alone, marking the best monthly gain since 2001. The earnings announced on the 26th (local time) also exceeded expectations.

In a conference call following the earnings release, Matthew Murphy, Marvell's CEO, predicted, "AI revenue, which reached $200 million this year, will double to over $400 million next year," adding, "It will double again in the 2025 fiscal year."

Founded in 1995, Marvell is an American semiconductor company specializing in data storage, communications, and system-on-chip (SoC) design. The semiconductors it produces play a crucial role especially in data centers. It designs controllers for data storage devices such as SSDs and HDDs, as well as chips that handle data communication between various chips.

Among these, 'chip communication' is a very important function in AI processing. Typically, when a GPU accelerates AI model processing, a data storage space is needed to hold the massive model capacity. This role is fulfilled by various memory chips. Therefore, data communication between memory chips and GPUs must be made as fast and efficient as possible for AI to operate smoothly.

Marvell produces data processing units (DPUs), chips that focus on this role, and supplies them to data centers. These chips not only facilitate smooth data processing but also reduce the burden on the central processing unit (CPU), the 'engine' of the data center, thereby lowering overall server power consumption. Considering the current concerns about environmental pollution due to power consumption in AI and data centers worldwide, these chips play a core role comparable to GPUs and CPUs.

The Last Survivor Among the Data Center Solution Trio That Dominated the 2010s

Marvell Technology U.S. Headquarters. [Image source=Marvell]

Marvell is not the only company making such DPUs. In fact, when Marvell emerged as a data center storage and communication solution company, there were already many rival companies. Among the most notable were 'Mellanox' and 'Xilinx.'

These three companies rose as representative growth firms during the explosive growth of the data center market in the 2010s, but now only Marvell remains. This is also due to the swift mergers and acquisitions by big tech companies that recognized the potential of AI semiconductor centers early on. Mellanox was sold to Nvidia in 2020 for $7 billion (approximately 9 trillion KRW), and Xilinx was merged into AMD as a subsidiary for a staggering $49 billion (approximately 64 trillion KRW).

Since then, Nvidia has fully utilized Mellanox's technology to develop its own DPU and GPU-memory chip connection solutions, currently reigning as the absolute leader in the AI semiconductor market. AMD is also preparing to stand out in data centers through collaboration with Xilinx.

However, these two companies are already large corporations exposed to the AI boom for some time and have diverse business areas. Currently, Marvell is the only company that can be evaluated purely as a 'data communication solution' provider.

As communication between GPUs and memory chips emerges as a key variable in AI computing, the number of companies producing DPUs is increasing. For example, Intel is also developing its own DPU solution called the 'Infrastructure Processing Unit (IPU).' With DPUs emerging as the 'third chip' constituting data centers, competition is intensifying.