Synopsys NPX6 Expands AI Options
The new deep-learning accelerator (DLA) can scale to more than 2,000 TOPS, providing a licensable core for applications such as autonomous driving and natural-language processing.
The DesignWare ARC NPX6 extends the Synopsys line of deep-learning accelerators (DLAs) with greater performance and flexibility to handle a broader range of neural networks. In its maximum configuration, it can deliver peak AI performance of more than 2,000 trillion INT8 operations per second (TOPS), plenty for high-end applications such as autonomous driving and natural-language processing. The company is licensing the NPX6 to lead customers and expects general RTL availability in 3Q22. The DLA is also available with a suite of safety features for ASIL B and D systems.
Each NPX6 core can generate 11 TOPS (without sparsity) from its multiply-accumulate (MAC) array, and a single cluster can scale to 24 cores. To achieve maximum performance, Synopsys assumes a design with eight full clusters. Relative to the previous EV7x core, the NPX6 adds optional floating-point support for customers that don’t want to quantize their models to INT8. It also has a flexible vector unit that can implement activation functions for a wide variety of neural networks, as disclosed at the recent Linley Processor Conference. It can optionally pair with a Synopsys VPX5 DSP for even greater flexibility.
The company has revised its product naming since releasing the EV7x two years ago. Whereas the EV name indicated a focus on embedded vision, the NPX name connotes a broader target of neural-network processing. Also, while the EV line started with the EV5x, Synopsys now considers that product the first generation, making the new design the sixth generation. The NPX6 competes against many other licensable DLAs but primarily Cadence’s DNA and Ceva’s NeuPro.