Calculations can be pooled, and memory can also be pooled. This is what MRVL's CEO said at the Taipei Computer Show in early June. The recently launched Structera CXL series by Marvel, including the Structera X memory expansion controller and Structera A near memory accelerator, is designed to address this issue. 1. Structera's CDB is purely hardware level, running line rate compression and decompression with full memory bandwidth (such as DDR5-6400 supporting 4 channels). This is completely transparent to the host CPU and operating system. The server "thinks" that it is connected to an 8TB memory block, but in reality, the physical memory is only 4TB. The original intention of CXL (Compute Express Link) was to decouple CPU and memory and establish a memory pool. But the biggest cost in CXL memory pools is still the expensive DRAM chips. Marvel uses compression technology to directly double the available memory capacity without changing the server, modifying the application, or adding additional DIMM memory modules. 2. This still holds great significance for Mrvl 1) Marvel is already an absolute leader in data transmission (optical communication DSP, network switch, PCIe Retimer), and Structera CXL has successfully shifted its business landscape from "data handling" to the core area of "data storage and memory expansion". 2) Memory accounts for a huge proportion of the cost of modern AI servers. Marvel's technology allows cloud vendors to double the effective memory capacity by purchasing the same amount of DRAM physical chips. This kind of underlying hardware that directly saves customers real money and silver gives Marvel extremely high product pricing power and gross profit margin space. 3. Boosting MRVL's future revenue 1) According to a report by Strategic Market Research, the global CXL component market (including controllers, switch chips, memory expanders, etc.) is expected to reach approximately $1.9 billion in 2024, and is projected to soar to $12.3 billion by 2030 (with a compound annual growth rate of over 32%) driven by the demand for AI heterogeneous computing. By 2035, the total market value of CXL controller IP and related components alone may approach $8 billion. 2) As a top player with exclusive "hardware compression" killer, Marvel is highly likely to cut 20% to 30% of its core market share in the future CXL controller and switch market. This means that in the medium to long term (2027-2030), the Structera series and subsequent CXL product lines are expected to bring Marvel pure incremental revenue of $1 billion to $3 billion annually. This not only fills the market gap, but also creates the strongest "second growth curve" for its annual revenue of $20 billion. 4. But it is difficult to completely break through the "memory wall" in the current computer architecture 1) Strictly speaking, the hardware compression of Marvel CXL, SanDisk's HBF flash memory, and even Nvidia and TSMC's HBM and CoWoS packaging are just "easing" or even "compromising", rather than truly "breaking through" the memory wall. 2) One of the core cornerstones of the von Neumann architecture is the physical separation of computing (CPU/GPU) and storage (Memory). In this architecture, all data must be transferred from memory to the computing unit through a channel called the "Bus", and then returned to memory after calculation. The essence of the 'memory wall' is the speed limit of this channel. No matter how wide the channel is widened (HBM), how close the external warehouse is pulled (HBF), or how tightly the goods are packed (Marvel compression), as long as the action of "moving" exists, the wall will always exist. The extremely high latency and huge energy consumption caused by data handling (in AI chips, the energy consumption of handling data is often more than a hundred times that of computing itself) are determined by physical laws. The only way to completely avoid hitting walls is to not move data. This requires us to thoroughly reconstruct the underlying operational logic of the computer. To completely break through the core of the "memory wall", one must first break through the "von Neumann" architecture of the computer system, which still seems a bit distant at present.