By Craig Barner, Senior Distinguished Engineer, Custom Silicon Engineering, Marvell
AI and confidential-compute workloads are scaling cryptographic demand faster than centralized HSMs (hardware security modules) can handle. With a centralized HSM, every operation you want to do has to be a network roundtrip from where your workload is to the centralized HSM and back. This process causes latency and scaling bottlenecks, and even an expanded attack surface, making systems and their data more vulnerable to hacking and leaking.
Under pressure to meet performance targets, teams start to compromise and put in workarounds. These workarounds—not HSMs—erode trust and security. A CVM (confidential VM) enclave where certificates and keys live in plain text in enclaves is vulnerable. VM Memory encryption does not mean that keys remain encrypted during application use within the VM. A dedicated HSM is needed.
Often in existing systems, critical keys are being generated outside of the centralized HSM boundary for convenience. They want to be where the workload is, and the same applies for keys being exported into environments that have weaker assurances. Another workaround to optimize performance is introducing local key caches, but again that erodes security by bypassing some of the security and isolation that the HSM offers. Keys are moving around different environments and different vendors, going from one cloud to another, often causing opportunities for keys to leak. The HSM is a secure fortress that provides an extra layer of security, but that is also causing scaling bottlenecks.
By Hsin-wei (Mark) Tseng, Director of Product Management, Custom Cloud Solutions, Marvell
How much memory bandwidth should an AI accelerator have?
With each new generation of XPUs, the answer grows. High-bandwidth memory (HBM) now takes roughly 30% of the area of a typical accelerator die and, at 2026 memory prices, more than half of XPU cost.1 Every HBM generation adds bandwidth, and with it a wider interface, more power and more heat in the same package.
At the 2026 OCP Global Summit in San Jose, Marvell presented its view on why custom HBM benefits from being designed as a system. Moving logic into a custom base die can free silicon for compute, it also bring new challenging on a new heat source under the DRAM stack. Area, power and heat are linked, so they cannot be solved one at a time. Marvell focuses on co-designing the base die, the die-to-die (D2D) interface and the package thermals together, solving them as one system problem for custom HBM.
By Vienna Alexander, Marketing Content Professional, Marvell
![]()
The Fast Mode Awards 2026 has recognized Marvell Photonic Fabric™ technology as an AI Compute Gamechanger for Impact. As a winner in the Data Center Solution Provider category, the award honors Marvell as a technology provider revolutionizing the next generation of data center capabilities and addressing the growing demands of the digital economy.
Memory capacity, bandwidth and connectivity are critical elements to consider in scale-up networks. With growing KV caches driving unprecedented memory demands, traditional architecture needs to adapt to create more efficient systems where processors can better access and move data across memory resources.
The Photonic Fabric technology platform enables more tightly integrated, high-bandwidth and power-efficient solutions for data center customers by creating pod-level optical shared memory.
By Arifur Rahman, Director of Product Marketing, Custom Cloud Solutions, Marvell
AI and ML workloads are rewriting the rules of data center design. Memory demands are reaching an unprecedented level due to growing LLM sizes, expanding context windows and growing KV-cache requirements. Traditional servers are not architected to meet this surge. Memory has become the primary bottleneck constraining AI performance and infrastructure efficiency. CXL-based memory expansion has emerged as one of the definitive architectural answers, and Marvell is raising the bar.
Building on the strong market adoption of the first-generation Structera® X CXL controllers, the second-generation Structera X 3406 and Structera X 3506, co-developed and co-architected with leading hyperscalers, address the most demanding AI infrastructure requirements.
Modern AI inference workloads, from LLM serving to deep learning recommendation models (DLRM), are memory bound. KV-cache capacity determines time-to-first-token (TTFT), throughput and XPU utilization. When memory runs out, performance collapses.
By Rajagopal Krishnaswamy, Associate Vice President, Product Marketing, Marvell
The networking industry has fully embraced the benefits of open networking. Organizations building cloud infrastructure, AI clusters, enterprise data centers and edge networks want the flexibility to choose the hardware and software solutions that best meet their requirements. Software for Open Networking in the Cloud (SONiC) has emerged as a powerful foundation for achieving that flexibility, enabling customers to adopt an open, community-driven network operating system across diverse environments.
But deploying SONiC in production requires more than installing software on a switch.
As networks grow in scale and complexity, organizations need validated software, operational visibility, lifecycle management, automation, and enterprise-grade support. Running modern infrastructure successfully requires a complete operational framework that helps teams deploy, manage, troubleshoot and evolve their networks over time.
That reality is driving the next phase of SONiC adoption.