Semidynamics Highly Configurable OOO Vector Unit
summary
Fully RVV1.0 compliant and offered in configurable sizes (start at 4, options up to 32 vector cores), the unit runs fully out-of-order vector renaming with mask and gather support, a customer-selectable vector register length, and a protocol to attach private RTL. An 8-core configuration achieves ~50% peak at 24×24 matrices and ~99% by 400×400 for FP workloads, FP16 hits 50% at 62×62 and ~99% by ~600, claims ~40 bytes/cycle to DDR and a 58% normalized advantage vs competitors.
AI-written summary · about