Halcyon unveils accelerator it says cuts model-training costs in half

Halcyon Systems introduced a new training accelerator on Tuesday that the company says halves the cost of training large language models, primarily by reducing the time processors spend waiting on one another.

Chief executive Marcus Ellery said the gains come from a redesigned interconnect that allows thousands of chips to share memory with far less coordination overhead. “The compute was never the problem,” he said. “The waiting was.”

Two cloud providers have committed to deploy the chips next year. Independent benchmarks are not yet available, and analysts say the cost claim will depend heavily on the workloads customers actually run.