Research Finder
Find by Keyword
Marvell Builds Breakthrough 2nm Custom SRAM for Next-Gen AI Infrastructure Silicon
Marvell advances its custom technology platform by introducing innovative 2nm custom SRAM, engineered to enhance the performance of custom XPUs and devices driving cloud data centers and AI clusters.
Key Highlights
- Delivers breakthrough bandwidth per sq mm with up to 6 gigabits of high-speed memory.
- Broadens Marvell’s custom technology platform to accelerate the innovation of AI infrastructure performance and cost-efficiency.
- Reduces total die area by up to 15% and cuts on-chip memory standby power by up to 66%.
The News
Marvell Technology, Inc., a supplier of data infrastructure semiconductor solutions, expanded its custom technology platform with the launch of its breakthrough 2nm custom Static Random Access Memory (SRAM), designed to boost the performance of custom XPUs and devices powering cloud data centers and AI clusters. For more details, check out the Marvell press release.
Analyst Take
Marvell unveiled an industry breakthrough by pioneering its 2nm custom SRAM, aimed at optimizing performance and efficiency for next-generation AI infrastructure silicon. Specifically, Marvell custom SRAM integrates advanced custom circuitry and software with cutting-edge 2nm process technology and core SRAM, providing up to 6 gigabits of high-speed memory while significantly lowering power consumption and die area at equivalent density.
Marvell's custom SRAM boosts memory hierarchy performance in accelerated infrastructure. Previously, Marvell launched its CXL technology for custom silicon integration, enabling terabytes of additional memory and enhanced compute capacity for cloud servers. It also introduced custom HBM technology, which increases memory capacity by up to 33% while reducing space and power needs for dense high-bandwidth memory (HBM) stacks in XPUs.
From my viewpoint, Marvell's custom SRAM offers the industry's highest bandwidth per square millimeter, allowing chip designers to reclaim up to 15% of the total area in a 2nm design. This freed-up silicon space can be used to add more compute cores, increase memory capacity, reduce device size and cost, or achieve a customized balance of performance, power, application needs, or total cost of ownership (TCO) goals. Featuring an innovative data path and architecture, Marvell's custom SRAM consumes up to 66% less power than standard on-chip SRAM at the same density while supporting operating speeds of up to 3.75 GHz.
Marvell Seizes Exploding Custom Silicon Opportunity
Marvell is demonstrating that the shift to custom solutions is rapidly growing and broadening. Once a speculative endeavor five years ago, custom solutions are now among the fastest-growing segments in the datacenter, driven by specialized workloads and improved cost and performance.
Marvell's 2nm custom SRAM capitalizes on the growing custom AI silicon trend by addressing the escalating demand for high-performance, power-efficient, and area-optimized solutions in AI-driven data center and edge applications, driven by hyperscalers and emerging customers seeking specialized chips like XPUs and XPU-attached solutions.
I see Marvell directly capitalizing on the custom chip market's total addressable market (TAM) surge from $42 billion to $55 billion in a year, with the overall data center equipment TAM also increasing. The customization of chips is expanding, with Marvell deftly introducing a new "XPU attached" segment. This category includes technologies such as PCIe retimers, NICs, CXL controllers, and other backend solutions designed to optimize XPU performance.
At its recent Custom AI Investor Event, Marvell shared its perspective that the market for XPU support chips (or custom XPU attach) is projected to grow dramatically, from $0.6 billion in 2023 to $14 billion by 2028 (based on combined Marvell, 650 Group and Dell’oro estimates). Notably, XPU support chips are outpacing XPUs in growth, with a 90% CAGR compared to 47% for XPUs. Marvell highlights a significant ratio of companion XPU support sockets, estimating two-thirds of sockets are for XPU support versus one-third for XPUs, underscoring the rapid rise of this segment.
Drill Down: Marvell Delivering a Cornerstone for the Custom Silicon Age
The customer base for custom chips is also diversifying, moving beyond the dominant "Big Four" hyperscalers. Marvell reports 18 active custom engagements, with 12 involving hyperscalers (three for XPUs, nine for XPU attach sockets) and four with emerging customers (two for XPUs, four for XPU attach sockets). Additionally, Marvell’s pipeline includes over 50 opportunities across more than 10 customers, with one-third focused on XPUs. The rise of emerging customers signals a broader adoption of custom chip solutions, highlighting the growing demand and opportunity in this evolving market.
As a result, Marvell’s custom SRAM represents the newest addition to its custom technology platform, tailored for chip development in a post-Moore’s Law era. For more than five decades, semiconductor performance and power improvements relied heavily on transistor scaling. However, as the costs and complexities of shrinking transistors escalate, semiconductor firms and their clients are shifting toward custom silicon solutions. These solutions are designed to align with specific customer infrastructure needs, incorporate innovative chip and packaging advancements, and challenge traditional chip design assumptions to push the limits of computing performance.
Drill Down: Marvell’s Unfolding Custom Platform Strategy
Marvell’s custom platform strategy is focused on achieving competitively advantageous outcomes through distinctive semiconductor designs and innovative methodologies. By integrating deep expertise in system and semiconductor design with advanced manufacturing processes, Marvell delivers a robust portfolio of platform solutions and intellectual property. This includes cutting-edge technologies such as electrical and optical serializer/deserializers (SerDes), die-to-die interconnects for 2D and 3D devices, silicon photonics, co-packaged copper, custom HBM, system-on-chip (SoC) fabrics, optical IO, and compute fabric interfaces like PCIe Gen 7.
Through close collaboration with customers, Marvell creates tailored platforms that significantly enhance infrastructure performance, efficiency, and value. I see this approach as enabling the development of highly optimized solutions that address the unique demands of next-generation AI and computing infrastructure, positioning Marvell as a leader in the custom silicon era.
Looking Ahead
Overall, I believe the future of AI infrastructure lies in customization. The advanced methodologies and technologies currently employed by hyperscalers to design state-of-the-art custom XPUs are poised to expand, reaching a broader range of customers, diverse device categories, and a wider array of applications. This shift signals a new era where tailored solutions drive innovation and performance across the AI landscape.
From my perspective, Marvell is most adept and passionate to partner with customers and collaborators to build a premier technology portfolio for this custom silicon age. By leveraging our expertise and fostering strong alliances, Marvell can deliver cutting-edge solutions that empower the next generation of AI infrastructure, meeting the evolving demands of this dynamic industry.
Ron Westfall | Analyst In Residence
Ron Westfall is a prominent analyst figure in technology and business transformation. Recognized as a Top 20 Analyst by AR Insights and a Tech Target contributor, his insights are featured in major media such as CNBC, Schwab Network, and NMG Media.
His expertise covers transformative fields such as Hybrid Cloud, AI Networking, Security Infrastructure, Edge Cloud Computing, Wireline/Wireless Connectivity, and 5G-IoT. Ron bridges the gap between C-suite strategic goals and the practical needs of end users and partners, driving technology ROI for leading organizations.