TAGS:

Marvell’s Silicon Strategy: Optimized Components For Different Workloads

Kurt Marko

Marvell has been rapidly building itself into a diversified supplier of IT infrastructure components. Through a combination of organic growth and recent acquisitions, Marvell has expanded its quarterly revenue by almost 70 percent over the past two years to more than $3.9 billion over the last four quarters.

2021 sales were greatly aided by two significant acquisitions: Inphi, which made Marvell a leader in silicon photonics; and Innovium, which gave it a high-performance Ethernet switching technology that targets hyperscale cloud operators. Concurrently, Marvell was first to market with a DPU that combined Arm’s v9 processor cores with TSMC’s 5nm fabrication technology, and is already porting its IP to 3nm and TSMC’s 2.5D Chip-on-Wafer-on-Substrate (CoWoS) packaging technology.

These developments were the foundation for Marvell’s recent analyst day event where the company detailed its strategy and vision for an era of cloud- and workload-optimized silicon.

Source: Marvell

As the requirements of cloud, edge, wireless carrier, automotive compute, and other markets diverge, Marvell seeks to tread a line between the cost-efficiency of merchant commodity products and the performance of semi-custom components built from a basket of standard modules. As Marvell President and CEO Matt Murphy said on its recent Q3 2022 earnings call,

“Similar to the rise of optimized silicon and cloud, automotive OEMs are realizing that to differentiate their products and deliver the most value to their customers, they need unique technology and IP to be embedded in compute silicon optimized to their specific platforms.”

The company target four market segments: data center systems, enterprise networking, wireless carrier infrastructure, and automotive. Within cloud data centers, Marvell further stratifies its products into five areas:

  • Compute: OCTEON, OCTEON Fusion (wireless baseband processing), and custom processors all using Arm cores.
  • Storage: Bravera SDD and HDD controllers
  • Networking: Teralynx (Innovium, maximum performance hyperscale switch), Prestera (mainstream enterprise switch), Alaska (Ethernet PHY)
  • Electro-optics: Pluggable and co-packaged optics with PAM4 DSP
  • Security: OCTEON, NITROX (cryptographic offload engines)

Source: Marvell

Innovation One: Electro-Optics And Co-Packaged Optics

Innovium gives Marvell a viable competitor to Broadcom in the cloud Ethernet switch market. In the DPU market, Marvell’s OCTEON-10 goes toe-to-toe with NVIDIA’s Bluefield-3. However, the two areas showing the most innovative differentiation for Marvell are electro-optics and wireless edge infrastructure. Indeed, Loi Nguyen, co-founder of Inphi and now EVP of Marvell’s optical and copper connectivity business, was a highlight of the day.

Nguyen began by nicely simplifying the three parameters used to scale optical or copper transmission bandwidth: the baud rate, modulation algorithm, and number of transmission lanes. Boosting speeds from 100G to 800G requires doubling each one by increasing the SerDes rate to 50G Gbit/s, the number of lanes from four to eight, and the modulation scheme from NRZ (a binary signaling method) to PAM4 (four-level pulse amplitude modulation). These elements are integrated into optical modules comprised of a PAM4 DSP, transimpedence amplifier (TIA; optical-to-electrical interface), and laser driver (electrical-to-optical interface). These modules can support a variety of optical interfaces; for example, 400GBASE-FR8 (short-haul Ethernet up to 2km) and 400ZR (for data center interconnect up to 80km using coherent DWDM).

Source: Marvell

Inphi was the first to offer integrated 400ZR and 400ZR+ (up to 400km) pluggable coherent optics modules that combine a 8-lane 56G PAM4 DSP, silicon photonics driver, and TIA and tunable laser. Soon these components will be integrated on a heterogeneous package using 2.5D Chip-on-Wafer-on-Substrate (CoWoS) packaging. Such multi-chip co-packages will eventually allow combining co-packaged optics (CPO) modules with switch silicon to increase interconnect density and bandwidth at significantly reduced power consumption.

Source: Marvell

Innovation Two: 5G Bringing Virtualization To The RF Edge

5G is catalyzing dramatic changes in the hardware architecture of wireless carriers as they emulate enterprises and cloud operators by using virtualization to disaggregate functions from monolithic proprietary hardware onto software modules running on programmable processors. At the extreme, the trend entails using vRAN to run 5G radio access networks on SoCs and splitting base station processing into a DU-CU (distributed-centralized unit) design.

Much like SDN and workload containerization, such software-based designs provide the flexibility to scale capacity and, as Vapor IO is doing with its INZONE edge services,  run other network services on servers at the border between metropolitan fiber plants and 5G wireless networks.

Source: Marvell

Marvell offers several SKUs of the OCTEON Fusion baseband DPUs, which are designed for vRAN RU (radio unit), DU, and CU deployments. These combine 12 to 36 64 Arm v8 cores with a security engine (NITROX V) that includes crypto acceleration, a DPI engine, hardware compression, and QoS on a SoC. Equipment vendors can also use Marvell’s cell library to design a custom SoC, such as one Samsung built for its massive MIMO radios that increase 5G capacity and coverage while consuming significantly less power than previous solutions using discrete components.

Marvell also just introduced the AtlasOne chipset optimized for 5G front haul networks that is the first to combine a 50Gbps PAM4 optical DSP with a TIA and choice of laser drivers supporting up to 40km link distances. When used with a vRAN architecture, AtlasOne and OCTEON Fusion allow carriers to scale radio capacity via support for up to 200 MHz bandwidth and 64 MIMO channels per RU.

Source: Marvell

Workload-Optimized Silicon Key To Cloud, Edge Expansion

By pursuing a parallel strategy of workload-optimized standard cells that can be mix-and-matched into high-volume merchant silicon or custom SoCs, Marvell is well-positioned to capitalize on the exploding requirement of cloud, 5G, and automotive industries for greater processing capacity, network throughput, and power efficiency.

Although the company’s recent analyst event was light on product announcements, it was outstanding at tying the many technologies Marvell has acquired and internally developed over the past year into a cohesive strategy. Its goal of providing integrated components targeting various workloads in cloud data centers, wireless and long-haul networks, and connected cars will significantly enhance the capabilities available to engineers in these disciplines.

About Kurt Marko: Kurt was an IT analyst, consultant and regular contributor to a number of technology publications including Diginomica, TechTarget and AvidThink. Starting his career as an electrical engineer, Kurt spent the past 35 years providing deep reporting and analysis in networking and IT. Kurt passed away in January 2022.