Marvell unveils 64 Gbps bi-directional die-to-die interface IP in 2 nm for next-generation XPUs

Marvell introduced a 2 nm 64 Gbps bi-directional die-to-die interface IP that delivers 32 Gbps of simultaneous two-way connectivity per wire to boost XPU bandwidth while reducing power and die area. The IP is also available in 3 nm and includes adaptive power management to cut interface power consumption.

Marvell Technology announced what it describes as the industry´s first 2 nm 64 Gbps bi-directional die-to-die (D2D) interconnect IP, aimed at improving bandwidth and performance for next-generation XPUs and data center designs. The interface provides 32 Gbps in each direction over a single physical wire, enabling simultaneous two-way connectivity. Marvell said the IP is also available in a 3 nm version and is intended to meet scaling demands in high-performance compute environments.

The company highlighted several technical metrics to quantify the new IP´s advantages. Marvell reported a bandwidth density greater than 30 Tbps per square millimeter, which it said is more than three times the bandwidth density of UCIe at equivalent speeds. A minimal depth configuration is claimed to reduce compute die area requirements to 15% compared to conventional implementations. The interface also incorporates advanced adaptive power management that automatically adjusts device activity to respond to bursty data center traffic, reducing interface power consumption by up to 75% with normal workloads and up to 42% during peak traffic periods.

Marvell positioned the D2D IP as setting a new standard for performance, power efficiency, and resiliency for chip designers targeting XPUs and next-generation data centers. The announcement focuses on technical gains in bandwidth density, silicon area savings and dynamic power reduction as primary benefits for designers aiming to pack more bandwidth into smaller die areas. Availability timelines, pricing and partner or customer details were not stated. Not stated.

72

Impact Score

Samsung to supply half of NVIDIA’s SOCAMM2 modules in 2026

Hankyng reports Samsung Electronics has secured a deal to supply half of NVIDIA’s SOCAMM2 modules in 2026 for the Vera Rubin Superchip, which pairs two ‘Rubin’ Artificial Intelligence GPUs with one ‘Vera’ CPU and moves from hardwired memory to DDR5 SOCAMM2 modules.

NVIDIA announces CUDA Tile in CUDA 13.1

CUDA 13.1 introduces CUDA Tile, a virtual instruction set for tile-based parallel programming that raises the programming abstraction above SIMT and abstracts tensor cores to support current and future tensor core architectures. The change targets workloads including Artificial Intelligence where tensors are a fundamental data type.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.