💢💢💢

🚀 Why might $Marvell Tech(MRVL.US)'s Structera become the most underestimated move in the era of AI infrastructure?

Marvell's CEO said something at this year's Computex that left a deep impression on me:

Computing can be pooled, and memory can also be pooled.

Behind this statement lies the core logic of Marvell's newly launched Structera CXL series (Structera X memory expansion controller + Structera A near-memory accelerator).

The biggest innovation of Structera is not CXL itself, but the built-in CDB (Compression Data Block) hardware compression engine.

It can perform line-speed compression and decompression at full DDR5 bandwidth, completely transparent to the CPU, operating system, and applications.

The server thinks it is connected to 8TB of memory, but in reality, there might only be 4TB of physical DRAM. Under workloads suitable for compression, the available memory capacity can nearly double, and this requires no modifications to server architecture, applications, or additional DIMMs.

What does this mean?

The original goal of CXL was to decouple the CPU and Memory, achieving Memory Pooling.

But the real cost has always been not CXL, but DRAM.

If the same amount of DRAM can provide higher effective capacity, cloud service providers can significantly reduce the total cost of ownership for AI servers, and that's precisely the value Marvell is providing.

More importantly, this is changing Marvell's strategic positioning.

In the past, Marvell was a key player responsible for "data movement" in AI infrastructure, including optical communication DSPs, network switching chips, PCIe Retimers, etc.

Now, Structera is pushing it further into the Memory Fabric and memory expansion domain.

In other words, Marvell is gradually moving from connecting data to managing data.

This is more important than simply launching a new chip.

From a market perspective, Strategic Market Research predicts that the global CXL component market will grow from about $1.9 billion in 2024 to about $12.3 billion in 2030, with a compound annual growth rate exceeding 30%.

If CXL memory pooling enters large-scale deployment in the future, and Marvell maintains its lead with hardware-level transparent compression, the Structera product line has the opportunity to become the company's second growth curve and contribute significant new revenue in the coming years.

However, I believe we must remain rational.

Many people say Marvell, HBM, HBF, and even CoWoS are "breaking through the memory wall." I don't fully agree.

More accurately, they are alleviating the memory wall, not eliminating it.

Because under the von Neumann architecture, computing and storage are inherently separate.

Data must constantly be moved between Memory and the CPU/GPU.

HBM widens the channels, HBF brings storage closer, and Marvell's hardware compression makes data smaller.

But as long as data still needs to be moved, latency and power consumption will not truly disappear.

To truly break through the memory wall, we must ultimately break through the traditional von Neumann architecture, making computing happen as much as possible where the data resides, rather than constantly moving the data.

This path is likely still a long way from true maturity.

What do you think will be the biggest bottleneck for future AI infrastructure: computing power, networking, or memory?

🔔 I will continue to share in-depth research on AI infrastructure, semiconductors, and long-term growth companies. Welcome to follow along as we search for the core winners in the next wave of technology.

#Marvell #MRVL #AIInfrastructure #CXL #Memory #Semiconductor #DataCenter #ArtificialIntelligence #CloudComputing #DDR5