Infrastructure Highlights from OCP Global Summit

Lots of hardware news came out of the OCP Global Summit, which the Open Compute Project held this week in San Jose, California, and some of it is relevant to our HPC audience. Here are the highlights of the announcements.

AMD introduced Helion, a rack design based on OCP specs that it plans to bring to market with partners like Cisco, Dell, HPE, Lenovo, Supermicro, and Wiwynn. AMD says it will work with the OEMs to outfit the 21-inch Helion racks (versus 19-inch for industry-standard racks) with its 6th Gen EPYC processors, as well as its new Instinct MI350P GPUs.

AMD will give partners and customers the flexibility to tailor CPU and GPU configurations, including liquid- and air-cooled versions. The company plans to begin shipping Helion by next month.

Not to be outdone, Intel shared more technical details on its upcoming D-Series GPU. Codenamed Crescent Island, the GPU is built on Intel’s Xe3P architecture and combines up to 480GB of LPDDR5X memory, 1.5TB/s of memory bandwidth, and 1.3 petaflops of FP8 compute, all within a 350-watt TDP.

The D-Series GPU, which is now sampling with customers, is designed to run the latest large language models and provide the memory bandwidth needed to manage the long context windows and concurrent demands of AI inference. Intel says the PCIe-based GPU will support up to 300 billion parameters on a single card, which will put it within range for LLMs with 1 trillion parameters in an eight card setup.

“Inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost,” stated Anil Nanduri, vice president of AI Products and GTM for the Intel Data Center Group. “With Intel GPU D-series, we’re focused on making high-capacity inference practical to deploy and economical to scale.”

d-Matrix unveiled Matrix Squadron, a fleet of multi-rack systems offering megawatts of compute capability based on its Corsair XPUs, JetStream NICs, and its Aviator software. Squadron is available in several configurations, scaling up to 16 d-Matrix SquadRacks, which are single eight-server racks, providing a cluster with up to 1,024 d-Matrix Corsair XPUs.

d-Matrix developed Squadron in partnership with Arista, Broadcom, and Supermicro. Squadron has already been deployed by d-Matrix in its Arizona data center for internal research and development, and is now shipping to priority customers for deployment in their own data centers.

“With Squadron shipping, and our Corsair servers and SquadRacks now generally available, the wait is over,” d-Matrix founder and CEO Sid Sheth stated. “Customers now have access to the world’s most efficient data center scale inference infrastructure available for building and ramping their ultra-low-latency AI applications.”

You can never have enough memory. In today’s age of massive KV caches, that saying has never been more true. With that in mind, Marvell Technology announced the second-generation updates to its Marvell Structera CXL memory expanders at OCP Global Summit.

The new Structera X 3406 supports up to 18 DDR4 RDIMM slots supporting 9TB of memory while the Structera X 3506 supports up to 12 DDR5 RDIMM slots supporting 6TB of memory in the CXL form factor.

“Marvell is the only provider with a complete end-to-end CXL portfolio spanning memory expansion (Structera X), near-memory acceleration (Structera A), and PFMA (Photonic Fabric Memory Appliance), pod-scale CXL memory-sharing appliance, giving customers a complete portfolio from direct-attach to disaggregated memory pooling,” stated Will Chu, executive vice president and general manager, Custom Cloud Solutions at Marvell. “We are delivering industry-leading CXL controllers, co-developed with leading hyperscalers to address the most demanding AI infrastructure requirements.”

Power is one of the big limiting factors in AI today, and OCP Global Summit featured various announcements on that front.

For staters, SolarEdge announced a solid-state transformer to support 800 Volt DC, the new standard that Nvidia is pushing to increase compute density in data centers. The company said that particular capability of the solid state transformer is currently running under load in its labs, with full power path and system-level validation expected soon.

“AI infrastructure is forcing the industry to rethink how power moves from the grid all the way to the rack,” stated Shuki Nir, Chief Executive Officer of SolarEdge. “Our work on MVAC-to-800 VDC power architecture is about applying SolarEdge’s deep experience in DC power electronics and protection to that challenge, while working with Nvidia and the broader ecosystem to help establish the common interfaces and protection principles needed to deploy these architectures at scale.”

SSD maker KIOXIA America announced new direct liquid cooling (DLC) models to its CM9 Series enterprise SSD lineup. The new E3.S DLC models support new cold-plate-enabled servers, as well as traditional air-cooled environments.

KIOXIA CM9-DLC Series SSD is compliant with PCIe 5.0, NVMe 2.0, and NVMe-MI 1.2c specs, as well as the new OCP Datacenter NVMe SSD v2.5 spec. It features capacities up to 30.72 TB, and offers sequential performance up to 14.8 GB/s read and 11 GB/s write.

“The rise of AI infrastructure is driving a fundamental shift in data center design toward higher power densities and advanced cooling technologies,” stated Maulik Sompura, a senior staff director with the company. “The KIOXIA CM9-DLC Series addresses these evolving requirements by combining industry-leading enterprise SSD performance with direct liquid cooling support in the E3.S form factor.”

Kioxia-CM9-DLC can use direct liquid cooling

Texas Instruments (TI) also made some news with several announcements, including the first analog-to-digital converter (ADC) geared toward enabling 24/7, low-power, event-based remote monitoring at nanosecond resolution.

TI says existing monitoring solutions “often fail to detect the nanosecond-level power-rail glitches that cause costly, unplanned downtime in data centers.” But thanks to the combination of ultra-high-speed sampling with advanced on-chip digital signal, the ADC34DSO75 can detect and capture transient fault events in real time.

TI also announced the LMI800PS, which is the industry’s “first 800V isolated hot-swap controller.” The new offering replaces a “complex discrete solution requiring five integrated circuits with a single high-density package that delivers in-rush current and power path protection.”

Finally, TI unveiled a new bulk acoustic wave (BAW)-based oscillator that delivers what it claims is best-in-class jitter performance and power-supply rejection ratio (PSRR). The LMK6B can detect jitters of 9.6fs at 625MHz, which TI says will ensure reliable performance and preserve signal integrity for AI network workloads.

Also on the network front, Micas Networks announced three new 102.4 Tbps open Ethernet switching platforms based on Broadcom Tomahawk 6 technology. The new gear includes:

Eliyan’s new Nexus and Oscar NuLight chiplets

Voyant Photonics used OCP Global Summit to reveal its plans to develop a chip-scale optical circuit switching (OCS) engine for AI data center networks. Voyant is designing the OCS Engine to be a compact, reliable optical switching building block that can be integrated into next-generation data center switching platforms.

“Rather than developing a complete networking system, Voyant plans to provide its OCS technology as an engine that optical networking companies can incorporate into their own platforms,” the company says.

Voyant has already industrialized the underlying silicon photonic switching architecture to meet the demanding requirements of LiDAR for physical AI, said Clement Nouvel, CEO of Voyant Photonics. “We are now applying that foundation to one of the most important emerging challenges in AI infrastructure,” he stated.

Eliyan launched a pair of new photonic products at the show designed to help customers blow past the copper wall. The new products are the Nexus and Oscar NuLight chiplets, which Eliyan labeled “the world’s first 2D microVCSEL interconnect platform.”

Nexus is a 1.6T reverse gearbox, while Oscar serves as an electrical interface chip (EIC) that connects high-speed electrical interfaces to microVCSEL and photodiode optical arrays. “Together, they provide scalable building blocks for linking compute resources across large AI systems, providing the enhanced reach, low power, and high reliability that these systems demand,” the company says.

Nvidia and Cisco launched a new pluggable optics form factor, dubbed Ultra-Dense Small Form Factor Pluggable (USFP), that the partners claim will help to scale AI factories by embracing optical networking technologies without overhauling their data center infrastructure.

“The first planned USFP-based transceivers will pack 16 full-duplex lanes in an octal small form factor (OSFP)-sized footprint supporting both switches and SuperNICs, with the ability to expand to 32 and more lanes in the future,” the companies say. “By doubling both the lane count and bandwidth per lane, USFP enables AI fabrics to connect more 1.6Tb/s endpoints without adding network tiers. Customers will be able to achieve 12.8Tb/s of overall capacity with USFP.”

Flex used the show to unveil new additions to its portfolio of data center racks dubbed Axiom AI Modules that adopt the latest in power and cooling technologies. These racks adopt 800 VDC power and advanced cooling distribution units (CDUs) to support the level of density that emerging AI workloads demands.

“The expanded portfolio is part of Flex’s Cloud and Power Infrastructure (CPI) segment, which Flex plans to separate into Axiom, an independent publicly traded company, in the first quarter of calendar 2027,” the company says.