Burlingame, CA15 days ago
Substantive working conversation, no prep, on: NPU/compute architecture (dataflow, memory hierarchy, MAC-array sizing, bandwidth-vs-compute balance); datatypes and quantization (INT4, INT8, FP8, BF16, OCP MX — numerics tradeoffs, not just names); SoC integration (AMBA AXI4/ACE-Lite, CoreSight, power-domain partitioning); functional safety (ISO 26262, FMEDA, lockstep, ASIL-B/D); and the competitive landscape (Synopsys NPX, Arm Ethos, Ceva NeuPro, VeriSilicon Vivante). Quadric's co-optimized software and hardware is targeted to run neural network (NN) inference workloads in a wide variety of edge and endpoint devices, ranging from battery operated smart-sensor systems to high-performance automotive or autonomous vehicle systems.