IBM Telum
IBM's 8-core mainframe microprocessor with on-chip AI inference, for the z16 series
Telum is a microprocessor designed by IBM and manufactured by Samsung Electronics for the IBM z16 series of mainframe computers. Announced at the Hot Chips 2021 conference on 23 August 2021, it was IBM's first processor to include on-chip acceleration for artificial intelligence inferencing while a transaction is running.
- Announced at Hot Chips 2021 on 23 August 2021
- 8 processor cores, 5.2 GHz clock
- 32 MB cache per core, scalable to 32 chips
- First IBM processor with on-chip AI inferencing during a transaction
- Does not thermal throttle; inserts sleep state instructions instead
What it is
Telum is a microprocessor built by IBM for the IBM z16 series of mainframe computers. IBM showed it at the Hot Chips 2021 conference on 23 August 2021, and Samsung Electronics manufactures it on a 7 nm process. The chip carries eight processor cores and runs at 5.2 GHz. What sets it apart is on-chip support for artificial intelligence inferencing happening at the same time as a business transaction, a first for an IBM processor. Mainframes are large, highly reliable computers used by banks, insurers and other organisations that process very large numbers of transactions.
How it came about
IBM announced Telum at the Hot Chips 2021 symposium on 23 August 2021. The design team faced the problem of running AI models inside transaction processing without sending data out to a separate accelerator. Their answer was to put inference support directly on the processor die. The chip went into the IBM z16 mainframe series, replacing the z15 generation. A follow-up part, Telum II, later added further functions to the same instruction set.
Inside the chip
The chip holds eight processor cores built around a deep superscalar out-of-order pipeline, meaning several instructions are worked on at once and can finish in a different order than they were issued. Each core has 32 MB of cache, the small fast memory sitting close to the core. The cache and interconnect scheme lets one core's L2 cache act as virtual L3 and L4 cache for another core, and the design scales to 32 Telum chips in one system. For AI work, Telum adds a 16-bit floating point number format called NNP-Data-Type-1 and the Neural Network Processing Assists instructions, which carry out tensor operations used by neural networks.
How it was made
Samsung Electronics fabricates Telum on a 7 nm process node. The die can be cooled with either air or water, though water cooling is needed once more than a few processors are packed into a single IBM compute drawer. Rather than lowering its clock speed when it gets hot, the chip inserts sleep state instructions instead of throttling.
Where it went
Telum serves as the processor for the IBM z16 series of mainframe computers. Those machines handle large volumes of business transactions, and the on-chip AI support lets a neural network model score a transaction while it is being processed. Systems can be built from up to 32 Telum chips, and larger configurations require water cooling.
Versions and family
Telum sits between two generations: it follows the z15 microprocessor and is followed by Telum II. Telum II extends the Neural Network Processing Assists instruction set with additional functions. The z/Architecture Principles of Operation manuals document the instruction set that these processors implement.
Why it matters
Telum marked the first time IBM put artificial intelligence inferencing acceleration directly on a processor so that it could run alongside a live transaction. Its cache arrangement, in which one core's L2 can serve as virtual L3 and L4 for another, offered a way to share fast memory across a large multi-chip mainframe system.
Questions people ask about IBM Telum
What is Telum?
Telum is a microprocessor designed by IBM and made by Samsung Electronics for the IBM z16 series of mainframe computers. It has eight cores running at 5.2 GHz and was announced on 23 August 2021 at the Hot Chips conference.
What makes it different from earlier mainframe chips?
It was IBM's first processor with on-chip acceleration for artificial intelligence inferencing while a transaction is taking place. It also introduces a 16-bit floating point format and the Neural Network Processing Assists instructions for tensor work in neural networks.
How is its cache organised?
Each of the eight cores gets 32 MB of cache. The design lets one core's L2 cache be used as virtual L3 and L4 cache for another core, and the interconnect scales to 32 Telum chips in a single system.
How is it cooled?
Telum can use either air or water cooling. Water cooling is required when more than a few Telum processors are installed in one IBM compute drawer. Instead of slowing its clock when hot, the chip inserts sleep state instructions.
What came before and after it?
Telum follows the z15 microprocessor and is succeeded by Telum II, which adds new functions to the Neural Network Processing Assists instructions.
Good reading and sources
- Checking your browserAnandTech, Article
- IBM’s New Telum Chip Reboots the MainframeIEEE Spectrum, Article
- The IBM mainframe: How it runs and why it survivesArs Technica, Article
The facts on this page come from the Wikipedia article "IBM Telum", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.