Research Finder
Find by Keyword
IBM z17: A Quantum Leap for Transactional AI
The IBM z17 aims to redefine AI at scale with its new Telum II processor, multi-model AI capabilities, and enhanced security features.
Key Highlights:
- IBM z17 is engineered for transactionalAI, powered by the new Telum II processor.
- The system aims to process 50% more AI inference operations per day than the z16.
- It introduces multi-model AI capabilities and new security features.
- The IBM Spyre™ Accelerator, expected in Q4 2025, will provide additional AI compute.
The News
IBM recently announced the z17, its newest mainframe system fully engineered for AI. Powered by the new IBM Telum II processor, the z17 expands the system's capabilities beyond transactional AI to enable new workloads. The IBM z17 will be generally available June 18, 2025, while the IBM Spyre Accelerator is expected to be available starting in Q4 2025. Find out more
What was Announced
Some specs of the newly announced box:
Box details
- 1-4 frames - all 19” rack with new doors that are cheaper to manufacture
- 208 customer cores.
- 11% improvement on single-thread performance vs z16
- 12-20% capacity growth Vs z16
- 60% more memory, up to 64TB of memory
- 17-27% less power than z16
Telum II Processor
- Samsung HPP process
- 43 billion transistors
- 24 miles of wire
- 18 layer wafer
- 8x 5.5Ghz cores per processor
- On-chip Data Processing Unit (DPU)
- 10x 36MB L2 cache
- 40% L2 Cache growth over z16
Spyre - PCIe connected
- 26 billion transistors
- 14 miles of wire
- 32 cores
- 2MB scratchpad
- PCIe gen5 x16 adapter
- 128GB of LPDDR5 memory
- 8 cards in I/O drawer form a logical cluster
- 1TB of memory
- 1.6TB per second aggregate memory bandwidth
The IBM z17 is architected to deliver improved AI inferencing capabilities, powered by a second-generation on-chip AI accelerator built into the IBM Telum II processor. I wrote about the Telum II at its announcement back at Hot Chips last year. Suffice to say it's a single-thread powerhouse, perfectly suited to transactional workloads. If AMD, Broadcom or Intel were to launch a 3Nm processor boasting a clock speed of 5.5Ghz, the internet would break with excitement and wonderment, but this is IBM and the mainframe, so analysts and a few hardcore fans like me will give it the kudos the IBM in-house, largely US based chip design deserves.
The system aims to enable more than 450 billion inferencing operations in a day with one millisecond response time. The IBM Spyre Accelerator, expected in 4Q 2025, aims to provide additional AI compute capabilities. At the analyst briefing a few weeks back, IBM was at pains to stress it has over 200+ use cases with early ship clients who are leveraging the AI capabilities of the box, and IBM went further to stress these use cases extend way beyond the obvious fraud mitigation example.
The system also includes new security features and tools that leverage AI for improved system usability and management. While continuing on a long established trend, the continued focus by IBM on post quantum encryption is to be applauded. In addition on the operations side, I am writing a white paper that is scheduled to publish in the next couple of weeks on IBM Z Operations Unite so you will have to wait for that to get my detailed perspective, suffice to say there is much to like with the combination of IBM's rich heritage with the OMEGAMON range of tools, combined with Open Telemetry and AIOps functionality.
What Does the New z17 Mean for IBM and the Industry as a Whole
This announcement needs to be seen in the context of the evolving AI landscape, the mainframe installed base, and IBM as a whole. Firstly, while cloud based AI solutions have gained significant traction, IBM is rightly doubling down on on-premises AI processing, aiming to deliver low latency and high throughput for critical applications. I fully expect to see big banks, telcos, retailers and government departments leveraging the AI capabilities of the Telum II and Spyre DPU (when it ships later in the year) to infuse transactional flows with AI. Secondly, the mainframe installed base. With this system coming 12 quarters after the z16, the upgrade financials will be more compelling for many of the installed base as they will have gone all the way through their leases, so IBM’s sales teams won’t need to perform the usual gymnastics to entice customers to upgrade.
This leads to the final point, the impact on IBM of the z17. The z17 hardware will drag 3-4x the revenue for the system, which I estimate to be in the $4bn range annually. The 3-4x drag will manifest in the likes of DS8000 storage, high margin transactional software such as CICS and Db2 and then support services and IBM Consulting. All told the long-stack revenue associated, will by my estimates, be in the region of $15bn, with the vast majority of it being at high margins. The IBM z17 matters to IBM revenue and margins.
With IBM already posting robust growth in the low single digit range, the next four quarters will have the tailwind of the mainframe refresh cycle, with the majority of clients most likely to upgrade in the first year of the cycle. Put simply, expect Arvind Krishna and Jim Kavanaugh to be talking about the mainframe refresh cycle in the coming quarters.
Analyst Take
We have had to wait longer than the typical mainframe refresh cycle (8-10 quarters) for the launch this week of the IBM z17. The IBM z16 was launched on the 5th April 2022 and has, according to the details shared at an Analyst pre-briefing a couple of weeks ago, been the most successful of the post-CMOS mainframe cycles. IBM only shares capacity growth by Millions of Instructions Per Second (MIPS) rather than systems or processors shipped. The current Telum II has 2,500 MIPS per processor for some context.
The current generation marks the 17th generation of CMOS systems since IBM moved from the previous bipolar systems in 1994. Given the innovation in the Telum II processor and accompanying Sypre PCIe-attached accelerator this extended cycle time is to be fully expected, and will no doubt lead to increased revenue in the early quarters of the new z17 cycle, more on that later.
The IBM z17 announcement signals a significant push towards integrating AI deeper into enterprise infrastructure. IBM's claim that the z17 is designed to process 50% more AI inference operations per day compared to the z16 is ambitious, but given my knowledge of IBM’s approach to the system claims, is to be believed. The introduction of multi-model AI capabilities suggests a move towards handling more complex AI workloads, potentially opening up new use cases for mainframe technology.
Looking Ahead
Based on what I am observing, the key trend that I am going to be looking out for is the adoption rate of the z17 and specifically the Spyre DPU beyond the obvious fraud use-case. Going forward I am going to be closely monitoring how the company performs on delivering on the real world impact of the Spyre accelerator. Customer adoption will be key.
When you look at the market as a whole, the announcement today highlights the ongoing importance of mainframes in handling demanding workloads, despite what the ill-informed will tell you about the platform. HyperFRAME will be tracking how the company does in future quarters and the impact of the mainframe on growth for the company as a whole. The key takeaway is that IBM is positioning the z17 as a platform designed for transactional AI and at scale, but its success will depend on its ability to deliver on its promises and meet the evolving needs of its customers.
Steven Dickens | CEO HyperFRAME Research
Regarded as a luminary at the intersection of technology and business transformation, Steven Dickens is the CEO and Principal Analyst at HyperFRAME Research.
Ranked consistently among the Top 10 Analysts by AR Insights and a contributor to Forbes, Steven's expert perspectives are sought after by tier one media outlets such as The Wall Street Journal and CNBC, and he is a regular on TV networks including the Schwab Network and Bloomberg.