IBM released its dataflow accelerator on October 13, 2026, promising a 2.3x speed boost for AI training workloads. It is the company's first major update to its AI hardware platform since the launch of its first generative AI chip in 2025.
IBM reported a 2.3x improvement in training throughput, measured on the standard AI benchmark for large language models. That compares with a 1.6x improvement from the company's previous architecture.
The dataflow accelerator is built on IBM's open agent factory framework and targets large-scale distributed inference tasks. Availability begins in the first quarter of 2027, initially for academic and research institutions.
"The dataflow accelerator is a key step in making AI-driven chip design more accessible," said Peter Hess, lead researcher at IBM. "This architecture enables efficient, scalable training for complex AI models."
The announcement follows IBM's recent collaboration with Coursera to expand AI education. IBM said the new accelerator will help scale AI training for educational and research purposes.
IBM did not say when the accelerator will be available for commercial use, and raised the question of how to balance performance with energy efficiency. The company said it will release further details in early 2027.
Source: ibm
