Trending Articles
Trending Topics
Trending
We’ve got AI in our phones, our laptops and our PCs. Now IBM is introducing them to its mainframes, with the all-new IBM z17. IBM claims this is the “first mainframe fully engineered for the AI age”.
This means that organisations large enough to run their own mainframes can execute a huge range of AI tasks on their own hardware. And not just GenAI tasks, but security tools that use AI to make it easier to spot threats and management tools for IT teams.
Here, we share the use cases the z17 mainframe is designed for, how it will tackle security risks associated with AI and the technology that powers it. Plus the small matter of when it goes on sale.
In a pre-briefing for the IBM z17 mainframe, Tina Tarquinio, Chief Product Officer for IBM Z and LinuxOne, was keen to emphasise that the platform is tailored to the needs of IBM customers.
“We’ve spent over 2,000 hours with almost 80 different companies across 36 different personas, across 16 different industries, to really understand how they want to use AI,” she explained.
“When and where they want to use AI, and what kind of outcomes do they really want to drive,” she clarified. “And they all have to do that while also serving their mission-critical applications.”
In the same briefing, Elpida Tzortzatos, IBM CTO AI on IBM Z and LinuxONE, was keen to point out how the IBM z17 mainframe met user needs.
She cited one example of embedding AI “in debit card transactions, in credit card transactions, in core payments, without slowing those transactions down”. The z17 will also help enterprises provide private GenAI for writing and summarisation, plus copilots to support developers.
Other use cases are coming, added Tzortzatos. “A new trend that we also see emerging with a lot of our clients that started with predictive AI around use cases for demand forecasting, for rescoring, for fraud, is the fact that now they’re combining both the strengths of predictive AI with the strengths of large language and CODA models to extract new features or new insights.”
One example: insurance claims. “We now see that a lot of the structured data around the policy of that insurance, the deductible, the amount of the claim, is in structured databases, such as Db2,” she said.
“And now we see clients deriving features from that structured data along with extracting key insights, let’s say from the description of a claim that’s unstructured text, to know the cause of the claim or the urgency of the claim. Then they use this enriched set of features that they fit into a predictive AI model to get better, more accurate results.”
Tzortzatos was also keen to point out how the IBM z17 mainframe met growing challenges over power consumption.
“Since traditional AI, when it comes to large language models and GenAI, we’ve seen a factor bigger than 100 in terms of model complexity and model size increase, and this leads to higher requirements for AI compute,” she said. That means higher power demands if you stick to older hardware.
“AI also brings new security challenges, and clients need to ensure that they can deploy their AI models in secure environments, or they risk exposing their data.”
Tzortzatos particularly pointed to a new threat posed by users crafting inputs that would “bypass restrictions to link to sensitive data or make the model behave unexpectedly”.
The z17 mainframe is built on IBM’s Telum II processor. Announced last August, this includes eight high-performance cores running at 5.5GHz. Compared to the Telum I, it includes 40% more cache.
But this release is all about AI, so IBM has boosted its AI throughput: “up to seven and a half times of the AI throughput that we had on z16”, said Tarquinio. The Telum II also has its own integrated AI accelerators, but it’s the new Spyre AI accelerators that really deliver.
“This is [an] IBM Research purpose-built engine with up to 32 cores on the Spyre accelerator, and that will be attached via PCIe card,” said Tarquinio. “And in the initial release we can have up to 48 of these Spyre accelerator cards on the system.”
She added: “It’s really exciting that we have this advanced capability with a focus on a decrease in energy. This system could do up to 450 billion inferences a day. Now, an inference is just one part of a transaction, but it’s really the part that could be the most powerful.
If you want an IBM z17 mainframe with Spyre Accelerators then you will have to wait: IBM says these are due for release Q4 2025.
However, the IBM z17 will available from 18 June 2025. IBM points potential buyers to its product website for more details.
