IBM Spyre
AI accelerator chip for IBM Z mainframes
IBM Spyre is an artificial intelligence accelerator designed by IBM Research for enterprise mainframes. It uses a 5 nm process and features 32 cores to handle generative AI workloads on premises.
- Contains 25.6 billion transistors on a 5 nm process
- Features 32 accelerator cores per chip
- Clusters of 8 cards provide 1 TB shared memory
- Supports int4, int8, fp8 and fp16 data types
- Designed for on-premises generative AI on mainframes
What it is
The IBM Spyre is a dedicated artificial intelligence accelerator chip. It is built to bring modern AI capabilities to IBM Z mainframe systems. IBM Research and IBM Infrastructure jointly developed the device. The chip allows enterprises to run generative AI models and perform inference directly on their secure hardware. This approach keeps sensitive data on premises rather than sending it to external cloud services. It represents a shift from general purpose computing to specialised matrix and vector processing for business applications.
How it came about
Development began with the Telum microprocessor introduced in 2022. Telum included a single AI accelerator core for real time transaction analysis. IBM Research then created a prototype unit with 32 cores to expand these capabilities. This prototype evolved into the enterprise grade Spyre product. The company previewed the final design at the Hot Chips conference in Palo Alto in 2024. By late 2025, IBM began shipping the accelerators as part of commercial bundles for its mainframe customers.
Inside the chip
Spyre contains 32 individual accelerator cores on a single die. It uses 25.6 billion transistors connected by 14 miles of wire. The architecture differs from standard central processing units by sending data directly between compute engines. This layout reduces energy use and memory demands for AI tasks. The chip supports lower precision numeric formats such as int4, int8, fp8 and fp16. These formats make running large language models more efficient. Each unit mounts on a peripheral component interconnect express card for easy installation.
How it was made
The chip is manufactured using 5 nanometre process technology. This advanced node allows for high transistor density and efficiency. Each accelerator is packaged onto a PCIe card. These cards are not sold individually. Instead, IBM offers them in eight card bundles. When clustered together, eight cards provide 256 accelerator cores. This configuration creates a virtual unit with one terabyte of shared memory and high bandwidth for complex model execution.
Where it went
Organisations use Spyre to deploy generative AI on mission critical mainframes. Common applications include automating business processes and modernising legacy code bases. The accelerator helps detect intricate fraud patterns in financial transactions. It works with IBM watsonx platforms to assist developers in updating application code. Banks and other enterprises benefit from keeping their data secure while accessing powerful AI inference tools. The system supports on premises model fine tuning and potential future training workloads.
Versions and family
The primary variant is the PCIe card form factor. IBM groups these into eight card clusters for higher performance. This setup aggregates resources to deliver over 2.4 petaops of throughput. The design shares architectural roots with the earlier Telum chip and the internal AIU prototype. While Telum had one core, Spyre scales this to 32 cores per chip. The company plans to integrate these accelerators into future generations of IBM Z systems.
Why it matters
Spyre enables the use of generative AI on systems that handle seventy percent of global transaction value. It addresses enterprise needs for secure, on premises AI deployment. By partnering with model builders like Anthropic, IBM aims to make these tools usable for legacy modernisation. The accelerator supports the transition from simple inference to more complex model tuning within corporate firewalls. This helps businesses maintain control over their data while adopting new artificial intelligence technologies.
Questions people ask about IBM Spyre
What is the IBM Spyre chip used for?
It is an AI accelerator for IBM Z mainframes. It allows enterprises to run generative AI models and perform inference securely on premises. Use cases include fraud detection and code modernisation.
How many cores does the Spyre chip have?
Each Spyre chip contains 32 individual accelerator cores. When installed in a cluster of eight PCIe cards, the system provides a total of 256 accelerator cores for heavy workloads.
Can you buy a single Spyre card?
No. IBM does not sell the PCIe cards separately. They are available only in eight-card bundles. These bundles are sold as full stack, value-priced packages for enterprise customers.
What manufacturing process is used for Spyre?
The chip is produced using 5 nanometre node process technology. This advanced manufacturing method allows for the integration of 25.6 billion transistors on a single die.
Good reading and sources
There is no English Wikipedia article on IBM Spyre yet, so the facts on this page come from the sources listed above: research.ibm.com/blog/spyre-for-z, nextplatform.com/ai/2025/10/10/ibm-ships-homegrown-spyre-accelerators-. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
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.