Home
Menu
Subscribe
Banner reading XPU dot pub
Metal-lidded chip with Marvell Octeon 10 superimposed

Marvell Octeon CN103 Keeps Comm Processors Alive


Marvell is preparing to ship production volumes of new, lower-cost, lower-power Octeon 10 SKUs. The Octeon 10 CN103 (and its lower-cost CN102 stablemate) features eight Arm Neoverse-N2 CPUs and 100-gigabit Ethernet ports as well as the packet processing engines that have characterized Octeon since its inception. Marvell calls Octeon a data-processing unit (DPU), a term more commonly used for smart-NIC chips, such as Nvidia BlueField or AMD Pensando chips than for more general processors like Octeon that target a range of communications applications. The CN103 can serve in smart NICs but mainly targets 5G small cells, SD-WAN systems as well as conventional firewalls and routers, and control-plane processing in systems like Ethernet switches.

Marvell Octeon CN103 Notables

  • CPU—Like other Octeon 10 models, the CN103 employs Arm’s Neoverse-N2, departing from past Octeon generations. As costly as the design may be to license, it’s less expensive and less risky than attempting to develop a better core in house. Marvell’s value comes from its other IP, the overall product definition, and software and support.
  • General performance—from a general-purpose-processing perspective, with two DDR5 ports, the CN103 is a midrange design.
  • Data-plane performance—the CN103 is a 60 Gbps design based on its IPSec throughput. The processor’s inline IPSec capabilities offload the CPUs. The CPUs aren’t needed for general network processing, a competitive advantage. Octeon 10 also implements vector packet processing, which groups similar packets so that long-latency operations such as lookup and decision trees occur once for the group instead of per packet, improving performance.
  • Nothing lower—The CN102 anchors the bottom of the Octeon 10 lineup. The market is now devoid of options for customers requiring a lower-performance, lower-cost communication (or even general embedded) processor. Available options lack the I/O throughput and whole-chip memory management and address translation.
  • Conservative investment era—In the past, new Octeon generations—each with multiple physical designs and each with faster high-end models—released often. The pace has slowed down. The first Octeon 10 model, the CN106 sampled in 2021. The CN103 is entering production three years later and is only the second (and likely the final) Octeon 10 chip (not counting the Octeon 10 Fusion processors). Processor suppliers have become much more conservative, rationing R&D and raising the revenue threshold for new products.

Competition

Most competing suppliers have effectively dropped out of the comm processor business, even declining to refresh products that generated hundreds of millions in revenue.

  • Broadcom bowed out, bogged down by trying to develop a high-performance CPU for the next generation of its XLP chips originating from its NetLogic acquisition and broaden its reach to target servers. Before its Marvell acquisition, Cavium bought the technology and rolled it into the Octeon line.
  • NXP hasn’t refreshed the Layerscape line, the Arm-based successor to PowerQuicc, having lost mobile-infrastructure opportunities (or not being invited to bid with the ill-fated Qualcomm acquisition pending), losing sockets to Intel, and finding it incongruous with the rest of the company.
  • Intel stands as Marvell’s lone competitor. A decade ago, the company had a record of consistent execution, it had narrowed the gulf in power and integration compared with embedded-/communications-processing competitors, and it had executive-level relationships with OEMs and service providers. Combined with backing initiatives such as SDN and NFV that promoted the idea that general CPUs could handle many networking functions and acquiring LSI’s (nee Agere’s) network-processing business, the company secured important wins.
    • The company developed the Xeon D line, which adds Ethernet and encryption acceleration to Xeon, and a series of Atom-based communications processors. These stand as Octeon’s main rival.
    • The CN103 offers more CPU performance than comparable Atoms and better packet-processing offloads than Xeon D. The newest Xeon Ds improve power efficiency over their predecessors, but we expect Marvell’s chips to have the advantage here.

Customers

  • Customers already using Octeon will find the CN103 offers better CPU performance and power at the same data-plane rates as earlier generations. Those already using an Octeon 10 and seeking to scale down their designs will also find the CN103 to be a good fit.
  • The CN103 also provides an Arm-compatible upgrade for customers using NXP Layerscape or Broadcom XLP processors, offering modern interfaces and bettering performance and power.
  • The most influential buyers for products like Octeon 10 are Cisco, Ericsson, and Nokia, and there is a long tail of customers needing products like the CN103. Marvell’s challenge is to provide an out-of-box experience comparable to that afforded by Xeon and support the distribution network.
  • Chinese companies such as Huawei and ZTE were once major drivers of comm-processing demand, but their influence has diminished, contributing to processor suppliers’ R&D conservatism.

Marvell Octeon CN103 Bottom Line

Through generations of refinement, Octeon maintains much of the original design philosophy as it updates its features. The CN103 is well-suited to enterprise networking gear, control-plane applications, and other systems requiring a moderate level of control- and data-plane processing. It also complements Marvell’s switch and PHY business, helping the company raise its bill-of-materials share of customers’ projects. Although Marvell has slowed its rate of processor development, it’s one of two competitors left. The CN103 differs sufficiently from Intel’s chips that Marvell will find many customer opportunities uncontested.


Posted

in

by


error: Unable to select