IBM expands mainframe networking options as AI reshapes enterprise infrastructure
Network World ·
IBM’s mainframe has come a long way from the time when it operated as a specialized networking island. Today, IBM Z is designed to be a high-performance participant in the wider enterprise network, using standard Ethernet and IP while adding sophisticated technologies for low-latency and high-throughput communications. IBM’s current mainframe networking strategy prioritizes support for conventional Ethernet/IP connectivity, increases remote direct memory access (RDMA)-based networking options, and enables proficient, secure communications with the large data center world that often spins around it. “The strategy is to ensure IBM Z is a deeply integrated part of a client’s broader hybrid infrastructure,” Andrew Crimmins , IBM master inventor and principal product manager for IBM Z, told Network World . “That means continuing to improve bandwidth, latency, and I/O capacity while making it easier for the mainframe to connect with cloud environments, distributed systems, storage, and AI infrastructure . DPU specialized for I/O acceleration One example of the mainframe’s evolution is the offloading of I/O processing onto purpose-built technology, such as the Data Processing Unit (DPU) architecture, so the system can handle growing network and storage demands more efficiently, Crimmins said. The DPU architecture is integrated with the IBM Telum II processor architecture that underpins current IBM Z systems. The DPU itself features four processing clusters, eight programmable microcontrollers per cluster, an I/O accelerator that coordinates the clusters and manages I/O subsystems, and a dedicated request manager, to ensure it can handle large I/O workloads, according to IBM . Offloading I/O function from the main processor cores isn’t a new strategy—IBM has done it in the past with custom ASICs on its I/O cards, Crimmins said. But the move to an on-chip DPU offers several benefits over the prior ASIC model, including: Clients continue to get the benefit of not spending Z core cycles on the overhead of I/O functions. The on-chip DPU has better performance, latency, and throughput. Removing many of the previously mentioned ASICs from the system (sometimes hundreds depending on the I/O card types and client requirements) saves power. Removing ASICs from the I/O cards allowed IBM to move to high-port-density cards for protocols like Fibre Channel/FICON, reducing the number of cards, which can reduce I/O drawer count, and even reduce the number of Z racks/frames in certain configurations, Crimmins noted. “We used to release those I/O ASICs about once every 10 years. They required a large additional investment, and we would end up building technical debt in the I/O space between generations,” Crimmins said. “By pulling that functionality up into the main processor chip, which we of course release a new version of with every mainframe generation, means we have a chance to get in there and innovate with every turn of the crank. It puts us in a much more agile position to deliver new technology faster to our clients.” Moving to 100GB Ethernet Another recent upgrade to the mainframe networking portfolio is the addition of a 100GB Ethernet adapter for the IBM Z and IBM LinuxONE systems. It features two 100 GbE ports with QSFP (Quad Small Form-factor Pluggable) SR optics that are plugged directly into the Central Processor Complex drawer, eliminating the need for external I/O drawers for high-speed network connectivity, according to IBM . Particularly with the LinuxONE mainframe, IBM urges customers to consolidate hundreds or thousands of Linux virtual machines onto a single mainframe, turning what might otherwise be a sprawling collection of distributed servers into a centralized network server. “Unlike your average ‘pizza box’ 1U server, IBM Z and LinuxONE are I/O heavy. In order to be the best at that, and handle it with unmatched resiliency so that our clients trust the platform with their most critical workloads, you need to be able to get a lot of data in and out at speed, with redundancy at every layer. When you have that much redundancy available, you are nowhere near as bandwidth constrained as a pizza box that may have one or two NICs,” Crimmins said. “100GbE was primarily requested by our clients who run Linux on the platform (LinuxONE or Linux on Z clients). Z and LinuxONE offer great sustainability benefits by enabling huge workload consolidations from 10,000+ x86 cores to a few hundred IFLs (Linux cores),” Crimmins said. He cited an example of Citibank , which migrated its “sprawling x86 installations” to a few LinuxONE servers, and “they got better performance, security and huge sustainability savings.” “Many other clients have seen that story and want to repeat the play with as many Linux workloads as possible,” Crimmins said. “We heard loud and clear from several clients that not having 100GbE prevented them from connecting into the portions of their org’s hybrid cloud environment that would allow them to compete for those workloads. So, we’ve delivered 100GbE to remove that blocker.” Another meaningful upgrade was Big Blue’s introduction of the Network Express converged adapter with the z17 family rollout in 2025. The converged adapter melds two previously separate networking architectures— OSA Ethernet and RMDA over converged Ethernet (RoCE/RDMA) as well as support for the Coupling Express 3 Long Reach function—onto the same physical adapter. The adapter can support 10G to 25G transmission speeds and link with single-mode fiber (LR/LX) and multi-mode fiber (SR/SX). The coupling function includes a two-port Ethernet-based, long-distance coupling card for tying together remote operations. Additions such as the Network Express converged adapter enable further physical IO consolidation as well as more flexibility for customers with less hardware, Crimmins said. That also means less of a hardware configuration burden and potentially fewer network connections to manage. In addition, Network Express uses what’s known as Enhanced Queued Direct I/O (EQDIO) to speed Ethernet traffic. The QDIO technology was designed for the Ethernet speeds and I/O architecture of earlier generations of Z mainframes. EQDIO redesigns that path to take advantage of the z17’s on-chip I/O processing and to handle much higher-throughput, lower-latency Ethernet traffic more efficiently. Taken together, enhancements like these further move the mainframe away from multiple specialized adapters and toward a converged Ethernet-based I/O architecture in which conventional IP networking and high-performance RDMA share the same physical infrastructure, Crimmins said. Keeping the mainframe relevant: Analyst reaction The networking advancements IBM has made are helping to keep the mainframe a first-tier participant in modern IT environments, said Steven Dickens , CEO and principal analyst with HyperFRAME Research . “IBM needs to keep evolving the mainframe in a variety of ways to keep it relevant, and it has done that repeatedly and successfully.” “Gone are the days when the mainframe was something weird and unusual. Now it is just another part of the modern networking environment,” Dickens said. As such, security is becoming more demanding. For example, segmenting and securing the network around the mainframe is something mainframe shops need to pay more attention to, Dickens said. “There was a time when businesses could do a security penetration test on a network tied to the mainframe maybe once every 12 months, but that certainly isn’t very strategic now,” Dickens said. Patching, too, is become an activity that needs to be kept up with more frequently, he said. IBM this summer rolled out the zSecure Detection package that combines threat monitoring, network insights, AI-driven access anomaly detection and automated response for z/OS. It includes automated network micro-segmentation capabilities that generate customized segmentation policies based on observed networking behavior. These policies can help organizations reduce unnecessary connectivity, limit opportunities for lateral movement, and strengthen security controls across critical environments, IBM stated. AI and the mainframe As AI infiltrates more technology environments, IBM sees a continuing role for the mainframe. “ Enterprise AI will rarely operate on one platform alone. Models, applications, and data will be distributed across mainframes, cloud platforms, accelerators and other infrastructure,” Crimmins said. “IBM Z’s differentiated role is its ability to apply AI close to the highly sensitive, high-volume transactional data already on the platform,” Crimmins said. “That can reduce unnecessary data movement while supporting faster decisions and stronger data governance.” “The mainframe can therefore remain a significant AI platform through a combination of differentiated on-platform AI capabilities, high-speed networking, and interoperability with the wider AI ecosystem,” Crimmins said. Beyond the network: Time-saving visualization and configuration tools Aside from networking-related advances, IBM’s Crimmins noted two other mainframe advances that are aimed at helping customers manage and grow their infrastructure. First is the IO Topology Visualization tool , which creates a graphical map of how an IBM Z system’s I/O infrastructure is physically and logically connected and shows connections among systems, adapters, fabrics, switches, and endpoints. The tool is part of IBM’s Infrastructure Management for Z and LinuxONE package. “We have a 50-page research summary where clients have shared example after example of diagrams of their environment they have painstakingly built by hand in tools like PowerPoint or Visio, showing the end-to-end mainframe – network (LAN/SAN) – storage and all the links between,” Crimmins said. “Customers told us it’s manual effort, painful, takes up time of the experienced folks on the team, and it’s always error prone at the end of the day. These diagrams are used for so many purposes though (planning to roll in new devices, troubleshooting, showing resiliency status to leadership to name a few) so there is no getting around creating them and keeping them updated. [We] just delivered the first release the IO Topology Visualization tool that does this for you, and we have a whole laundry list of features on roadmap to enhance it in the future.” The second new technology is the Infrastructure as Code via IBM Infrastructure Manager and IBM Terraform package. “While configuring mainframe infrastructure gives users very granular control, which is good, it has the downside of any specialized tooling—it requires specialized skills,” Crimmins said. “IBM Infrastructure Management for IBM Z and LinuxONE just recently became generally available. This is a modern interface that also allows it to pair with IBM Terraform to enable users to leverage broadly leverage Infrastructure as Code on our platform for the first time.” Infrastructure as code is common in the distributed computing space, and young people are coming out of universities and colleges with IaC skills, “which makes their time to proficiency on our platform drastically shorter than learning bespoke tools from scratch,” Crimmins said. “IaC allows for configuration of I/O resources and Logical Partitions, while also playing well with industry standard automation like leveraging Ansible playbooks. This a huge leap forward in how agile our clients can be in setting up environments, leveraging best practices via pre-written playbooks, and making things repeatable for situations like setting up identical environments in hours rather than weeks or months.”
IBM’s mainframe has come a long way from the time when it operated as a specialized networking island. Today, IBM Z is designed to be a high-performance participant in the wider enterprise network, using standard Ethernet and IP while adding sophisticated technologies for low-latency and high-throughput communications. IBM’s current mainframe networking strategy prioritizes support for conventional Ethernet/IP connectivity, increases remote direct memory access (RDMA)-based networking options, and enables proficient, secure communications with the large data center world that often spins around it. “The strategy is to ensure IBM Z is a deeply integrated part of a client’s broader hybrid infrastructure,” Andrew Crimmins , IBM master inventor and principal product manager for IBM Z, told Network World . “That means continuing to improve bandwidth, latency, and I/O capacity while making it easier for the mainframe to connect with cloud environments, distributed systems, storage, and AI infrastructure . DPU specialized for I/O acceleration One example of the mainframe’s evolution is the offloading of I/O processing onto purpose-built technology, such as the Data Processing Unit (DPU) architecture, so the system can handle growing network and storage demands more efficiently, Crimmins said. The DPU architecture is integrated with the IBM Telum II processor architecture that underpins current IBM Z systems. The DPU itself features four processing clusters, eight programmable microcontrollers per cluster, an I/O accelerator that coordinates the clusters and manages I/O subsystems, and a dedicated request manager, to ensure it can handle large I/O workloads, according to IBM . Offloading I/O function from the main processor cores isn’t a new strategy—IBM has done it in the past with custom ASICs on its I/O cards, Crimmins said. But the move to an on-chip DPU offers several benefits over the prior ASIC model, including: Clients continue to get the benefit of not spending Z core cycles on the overhead of I/O functions. The on-chip DPU has better performance, latency, and throughput. Removing many of the previously mentioned ASICs from the system (sometimes hundreds depending on the I/O card types and client requirements) saves power. Removing ASICs from the I/O cards allowed IBM to move to high-port-density cards for protocols like Fibre Channel/FICON, reducing the number of cards, which can reduce I/O drawer count, and even reduce the number of Z racks/frames in certain configurations, Crimmins noted. “We used to release those I/O ASICs about once every 10 years. They required a large additional investment, and we would end up building technical debt in the I/O space between generations,” Crimmins said. “By pulling that functionality up into the main processor chip, which we of course release a new version of with every mainframe generation, means we have a chance to get in there and innovate with every turn of the crank. It puts us in a much more agile position to deliver new technology faster to our clients.” Moving to 100GB Ethernet Another recent upgrade to the mainframe networking portfolio is the addition of a 100GB Ethernet adapter for the IBM Z and IBM LinuxONE systems. It features two 100 GbE ports with QSFP (Quad Small Form-factor Pluggable) SR optics that are plugged directly into the Central Processor Complex drawer, eliminating the need for external I/O drawers for high-speed network connectivity, according to IBM . Particularly with the LinuxONE mainframe, IBM urges customers to consolidate hundreds or thousands of Linux virtual machines onto a single mainframe, turning what might otherwise be a sprawling collection of distributed servers into a centralized network server. “Unlike your average ‘pizza box’ 1U server, IBM Z and LinuxONE are I/O heavy. In order to be the best at that, and handle it with unmatched resiliency so that our clients trust the platform with their most critical workloads, you need to be able to get a lot of data in and out at speed, with redundancy at every layer. When you have that much redundancy available, you are nowhere near as bandwidth constrained as a pizza box that may have one or two NICs,” Crimmins said. “100GbE was primarily requested by our clients who run Linux on the platform (LinuxONE or Linux on Z clients). Z and LinuxONE offer great sustainability benefits by enabling huge workload consolidations from 10,000+ x86 cores to a few hundred IFLs (Linux cores),” Crimmins said. He cited an example of Citibank , which migrated its “sprawling x86 installations” to a few LinuxONE servers, and “they got better performance, security and huge sustainability savings.” “Many other clients have seen that story and want to repeat the play with as many Linux workloads as possible,” Crimmins said. “We heard loud and clear from several clients that not having 100GbE prevented them from connecting into the portions of their org’s hybrid cloud environment that would allow them to compete for those workloads. So, we’ve delivered 100GbE to remove that blocker.” Another meaningful upgrade was Big Blue’s introduction of the Network Express converged adapter with the z17 family rollout in 2025. The converged adapter melds two previously separate networking architectures— OSA Ethernet and RMDA over converged Ethernet (RoCE/RDMA) as well as support for the Coupling Express 3 Long Reach function—onto the same physical adapter. The adapter can support 10G to 25G transmission speeds and link with single-mode fiber (LR/LX) and multi-mode fiber (SR/SX). The coupling function includes a two-port Ethernet-based, long-distance coupling card for tying together remote operations. Additions such as the Network Express converged adapter enable further physical IO consolidation as well as more flexibility for customers with less hardware, Crimmins said. That also means less of a hardware configuration burden and potentially fewer network connections to manage. In addition, Network Express uses what’s known as Enhanced Queued Direct I/O (EQDIO) to speed Ethernet traffic. The QDIO technology was designed for the Ethernet speeds and I/O architecture of earlier generations of Z mainframes. EQDIO redesigns that path to take advantage of the z17’s on-chip I/O processing and to handle much higher-throughput, lower-latency Ethernet traffic more efficiently. Taken together, enhancements like these further move the mainframe away from multiple specialized adapters and toward a converged Ethernet-based I/O architecture in which conventional IP networking and high-performance RDMA share the same physical infrastructure, Crimmins said. Keeping the mainframe relevant: Analyst reaction The networking advancements IBM has made are helping to keep the mainframe a first-tier participant in modern IT environments, said Steven Dickens , CEO and principal analyst with HyperFRAME Research . “IBM needs to keep evolving the mainframe in a variety of ways to keep it relevant, and it has done that repeatedly and successfully.” “Gone are the days when the mainframe was something weird and unusual. Now it is just another part of the modern networking environment,” Dickens said. As such, security is becoming more demanding. For example, segmenting and securing the network around the mainframe is something mainframe shops need to pay more attention to, Dickens said. “There was a time when businesses could do a security penetration test on a network tied to the mainframe maybe once every 12 months, but that certainly isn’t very strategic now,” Dickens said. Patching, too, is become an activity that needs to be kept up with more frequently, he said. IBM this summer rolled out the zSecure Detection package that combines threat monitoring, network insights, AI-driven access anomaly detection and automated response for z/OS. It includes automated network micro-segmentation capabilities that generate customized segmentation policies based on observed networking behavior. These policies can help organizations reduce unnecessary connectivity, limit opportunities for lateral movement, and strengthen security controls across critical environments, IBM stated. AI and the mainframe As AI infiltrates more technology environments, IBM sees a continuing role for the mainframe. “ Enterprise AI will rarely operate on one platform alone. Models, applications, and data will be distributed across mainframes, cloud platforms, accelerators and other infrastructure,” Crimmins said. “IBM Z’s differentiated role is its ability to apply AI close to the highly sensitive, high-volume transactional data already on the platform,” Crimmins said. “That can reduce unnecessary data movement while supporting faster decisions and stronger data governance.” “The mainframe can therefore remain a significant AI platform through a combination of differentiated on-platform AI capabilities, high-speed networking, and interoperability with the wider AI ecosystem,” Crimmins said. Beyond the network: Time-saving visualization and configuration tools Aside from networking-related advances, IBM’s Crimmins noted two other mainframe advances that are aimed at helping customers manage and grow their infrastructure. First is the IO Topology Visualization tool , which creates a graphical map of how an IBM Z system’s I/O infrastructure is physically and logically connected and shows connections among systems, adapters, fabrics, switches, and endpoints. The tool is part of IBM’s Infrastructure Management for Z and LinuxONE package. “We have a 50-page research summary where clients have shared example after example of diagrams of their environment they have painstakingly built by hand in tools like PowerPoint or Visio, showing the end-to-end mainframe – network (LAN/SAN) – storage and all the links between,” Crimmins said. “Customers told us it’s manual effort, painful, takes up time of the experienced folks on the team, and it’s always error prone at the end of the day. These diagrams are used for so many purposes though (planning to roll in new devices, troubleshooting, showing resiliency status to leadership to name a few) so there is no getting around creating them and keeping them updated. [We] just delivered the first release the IO Topology Visualization tool that does this for you, and we have a whole laundry list of features on roadmap to enhance it in the future.” The second new technology is the Infrastructure as Code via IBM Infrastructure Manager and IBM Terraform package. “While configuring mainframe infrastructure gives users very granular control, which is good, it has the downside of any specialized tooling—it requires specialized skills,” Crimmins said. “IBM Infrastructure Management for IBM Z and LinuxONE just recently became generally available. This is a modern interface that also allows it to pair with IBM Terraform to enable users to leverage broadly leverage Infrastructure as Code on our platform for the first time.” Infrastructure as code is common in the distributed computing space, and young people are coming out of universities and colleges with IaC skills, “which makes their time to proficiency on our platform drastically shorter than learning bespoke tools from scratch,” Crimmins said. “IaC allows for configuration of I/O resources and Logical Partitions, while also playing well with industry standard automation like leveraging Ansible playbooks. This a huge leap forward in how agile our clients can be in setting up environments, leveraging best practices via pre-written playbooks, and making things repeatable for situations like setting up identical environments in hours rather than weeks or months.”