←
Marvell Technology5,463 posts
Marvell Technology avatar

Marvell Technology

@MarvellTech

Essential technology, done right™ Data Center & Cloud | Carrier | Enterprise

◷ Joined August 2009▧ 2,188 Photos and videos
Marvell Technology avatar
Marvell Technology@MarvellTech·19h

What if AI memory could move beyond the boundaries of a single server? On display at Supercomputing 26: @WeAre_Penguin’s MemoryAI™ Photonic Memory Module, the world’s first silicon photonics memory module designed to deliver HBM-class performance, built on the Marvell® Photonic Fabric™ technology platform. The module is a building block for Penguin Solutions’ MemoryAI™ Photonic Appliance, designed to enable transparent sharing of disaggregated high-bandwidth memory across multiple CPU- and GPU-based servers. See the innovation taking shape at #SC26 in Chicago at Marvell booth 4003 or Penguin Solutions booth 1723.

Post attachment
Penguin Solutions avatar
⟳ Marvell Technology retweeted
Penguin Solutions@WeAre_Penguin·20h

Live from #SC26 in Chicago, Penguin Solutions is breaking the GPU memory wall for agentic AI. We’ve unveiled the MemoryAI™ Photonic Appliance: the industry’s first photonics-based HBM-class memory appliance, alongside its underlying MemoryAI™ Photonic Memory Module, built on the @MarvellTech Photonic Fabric™ technology platform. Today’s agentic AI workloads are hitting a brutal wall: GPUs can spend 70% of their time sitting idle, waiting for data transfers. The physical constraint is clear. Ultra-fast HBM historically had to sit mere millimeters from the accelerator die, severely constraining memory capacity to package size. By replacing conventional copper with silicon photonics, this industry-first appliance extends HBM-class data rates across meters rather than millimeters. GPUs and CPUs can now transparently pool and access memory across a multi-rack cluster with zero data copying required. Visiting #SC26? Stop by our booth #1723 to see it in action, and read the full product launch details here: ir.penguinsolutions.com/news…

Post attachment
Marvell Technology avatar
Marvell Technology@MarvellTech·20h

HBM now takes roughly 30% of an AI accelerator die and, at 2026 memory prices, more than half of its cost. Custom HBM can free up to 25% more die area for compute, depending on the XPU design. But moving logic into a custom base die also moves heat under the DRAM stack. Hsin-wei (Mark) Tseng, Director of Product Management for Custom Cloud Solutions at Marvell, explains why the base die, the die-to-die interface and the package thermals need to be designed together. The article also covers how the industry is tackling heat and how adaptive power management can cut interface power by up to 75% under normal workloads. Read the full article: mrvl.co/4tp1SF3

Post attachment
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 16

Marvell Photonic Fabric™ technology enables up to 32TB of warm KV cache offload across multiple XPUs and racks up to 50 meters apart. @TheFastMode Awards 2026 has recognized the technology as an AI Compute Gamechanger for Impact, naming Marvell a winner in the Data Center Solution Provider category. By loading KV cache from a shared memory tier instead of storage, Photonic Fabric helps increase inference throughput, support larger models and longer context lengths, and deliver up to 2 to 3x higher token throughput within existing data center footprints and power envelopes. Read more about the award and the technology: mrvl.co/4zu4YJg

Post attachment
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 16

AI inference is scaling. Memory needs to keep up. Marvell and @WeAre_Penguin, the AI Factory Platform Company, are collaborating to address one of AI infrastructure's growing challenges: memory capacity, bandwidth and latency. The collaboration brings together Marvell® Structera™ X memory-expansion controllers, designed to extend server memory capacity by reusing existing DDR4 memory, and Marvell® Photonic Fabric™ optical interconnect technology, designed to enable high-bandwidth, low-latency memory sharing across multiple racks. Together, these complementary technologies offer a roadmap from direct-attach memory expansion to disaggregated, multi-rack sharing, helping data centers scale memory as AI inference demands evolve. Discover how Marvell and Penguin Solutions are advancing memory expansion from the server to the pod: mrvl.co/4eH5iN3

Post attachment
Six Five Media avatar
⟳ Marvell Technology retweeted
Six Five Media@TheSixFiveMedia·Nov 13

“This era of connectivity is going to unleash a whole new wave of innovation.” @MarvellTech CEO Matt Murphy opened Day 3 of the Six Five Summit with @PatrickMoorhead and @danielnewmanUV, and shared how optics could cut data center power 25%+ and a push past the memory wall. $MRVL

1:22
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 13

Silicon photonics could break the bandwidth and power barrier facing AI and HPC systems. What will it take to deploy it at scale? Preet Virk, Senior Vice President and General Manager of the Photonic Fabric Business Unit at Marvell, will join a panel of experts at #SC26 to discuss the value of photonic interconnects today, the technical and operational challenges ahead and what the ecosystem must do to accelerate adoption. Learn more: mrvl.co/4zGKi1q

Post attachment
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 13

As AI and HPC workloads grow, more compute and memory-intensive infrastructure is encountering both connectivity and memory bottlenecks that are causing significant underutilization. At SC26, booth 4003, Marvell will showcase the technologies enabling the next generation of AI infrastructure across scale-up, scale-out and scale-across architectures, including technologies that scale memory and interconnect as a single system. Demonstrations include: ➡️Marvell® Photonic Fabric™ memory module ➡️CXL memory solutions with the Marvell Structera™ add-in card ➡️Multipath reliable connection (MRC) with Marvell Teralynx® 10 switch silicon ➡️Marvell Alaska® A AEC DSP-based 1.6T and 800G active electrical cables Preet Virk, Senior Vice President and General Manager of the Photonic Fabric Business Unit at Marvell, will join the Birds of a Feather panel “Photonics for AI and HPC Clusters: Breaking the Bandwidth and Power Barrier” on Wednesday, November 18. Read the press release: mrvl.co/4f9gdPN

Post attachment
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 11

Since 2022, the Technical Leadership Mentoring Program has reached 63 senior technical leaders at Marvell across four cohorts, supported by more than 30 executive mentors. The 12-month program, sponsored by Corporate Chief Technology Officer Noam Mizrahi, combines immersive workshops, a leadership 360 assessment and 1:1 mentoring with executives. Participants learn to shape strategy, lead cross-functional work and influence teams they don’t manage. Distinguished Engineer Dan Scofield says the program changed how he thinks about his role, “not just as a technical resource, but as a connector and enabler for others.” Learn how the program develops technical leaders: mrvl.co/4eYin4I

Post attachmentPost attachmentPost attachmentPost attachment
Six Five Media avatar
⟳ Marvell Technology retweeted
Six Five Media@TheSixFiveMedia·Nov 9

KV caches have grown about 10x in nine months, @MarvellTech's Dave Lazovsky tells @PatrickMoorhead and @danielnewmanUV. Holding that much model and cache in HBM means scaling past the rack, from 144 XPUs toward 576, and that makes optical scale-up networks a must-have. $MRVL

1:19
Six Five Media avatar
⟳ Marvell Technology retweeted
Six Five Media@TheSixFiveMedia·Nov 6

Marvell’s data center business now runs above $2 billion a quarter, and mass deployments of scaled-up AI accelerators haven’t happened yet. @MarvellTech’s CEO Matt Murphy tells @PatrickMoorhead and @danielnewmanUV why connectivity, from near-package optics to coherent links between data centers, is next. $MRVL

1:59
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 5

Ara T can cut optical module power consumption by more than 35%. It’s part of the Marvell 1.6T optical DSP platform portfolio, which just earned a 2026 LEAP (Leadership in Engineering Achievement) Award from @DesignWorld in the connectivity category. The portfolio includes Ara T, the first 8x200G transmit-retimed optics DSP; Ara X, the first 1.6T DSP with advanced link reliability; Petra, the first 3nm 8x100G to 4x200G gearbox; and Aquila M, the first O-band-optimized coherent-lite DSP with integrated MACsec. Read more about the award-winning portfolio: mrvl.co/4ekLX4c

Post attachment
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 4

Are you looking for your next engineering role? Marvell is attending the Society of Women Engineers WE26 conference. Engineers from Women@Marvell will join our recruiters at the career fair to meet candidates for internships and full-time roles. Visit booth 3037 to meet the team. Learn more about careers at Marvell: mrvl.co/48z843m #WE26

Post attachment
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 3

AI clusters run at just 40-50% GPU utilization today. The bottleneck has moved from compute to memory to connectivity. In a new video, Rajagopal Krishnaswamy, AVP of Product Marketing for Cloud Switch at Marvell, shows a live Teralynx T100 running traffic at full wire speed. The T100 is the industry’s first monolithic 102.4 terabit per second Ethernet switch built on 3nm. Its programmable pipeline, adaptive routing and congestion signaling keep traffic moving so GPUs keep working. It’s also open and standards-based, so existing tools and operations carry over. Learn more about the performance, connectivity and programmability of Teralynx T100 Ethernet switches: mrvl.co/4koQoyy

2:09
Marvell Technology avatar
Marvell Technology@MarvellTech·Nov 3

As AI infrastructure scales and optical connectivity moves closer to compute, deeper integration of semiconductor and optical technologies is becoming essential to delivering the bandwidth, performance and power efficiency customers need. Marvell Chairman and CEO Matt Murphy and @Lumentum President and CEO Michael Hurlston joined CNBC's @SquawkStreet to discuss how the combination of the companies will create the category leader in AI connectivity. Watch the interview: mrvl.co/4jUwfAs Important information and disclosures: mrvl.co/3WJphEB

Post attachment