Cadence Boosts DLA Speed By 20×

Cadence’s Neo AI licensable IP block provides up to 80 TOPS of performance. Its NeuroWeave SDK unifies AI design across all Cadence IP options.

Bryon Moyer

Cadence’s Neo licensable AI core supplants its older NNA 110 core with 20× higher peak performance. At 80 TOPS, it aims at the high end of the edge market. The company’s new NeuroWeave tools provide a unified AI design environment across all licensable Cadence cores targeting neural networks.

Neo’s performance originates in the core’s maximum configuration of 32K multiply-accumulate units. Area efficiency, compared with the older model, improved by as much as 4×; power efficiency rose by as much as 12× compared with the NNE 110—a smaller version of the NNA 110 that remains active for new designs. Prior versions required a separate Cadence DSP host, but Neo can operate with any host accessible over an AXI bus.

Eighty TOPS is a large number compared with prior typical edge solutions. Specific applications expected to require that level of performance include automotive for L3/L3+ designs with automotive safety-integrity level (ASIL) B safety certification (which Cadence has not yet acquired) and augmented/virtual reality. The company said many of the requests it gets are for 50 TOPS and above; the prospects aren’t sure what they’ll do with all that, but they’re future-proofing to ensure they have the required horsepower after future updates.

Cadence’s various licensable AI designs have required their own tools. The new NeuroWeave SDK (software-development kit) unifies design through one tool, allowing licensees to design the same way for cores and to easily switch among them.

Available today, both the new core and the new tools help rationalize what has been a fractured lineup. Cadence’s smallest units remain available and recommended for new designs; the new core extends the high end (replacing the NNA 110) while overlapping the larger DSP-based options. The new tool brings them together with a consistent design environment.