An 8x8x8 matrix multiplication hardware accelerator with a 64-MAC spatial array, 128 kB scratchpad memory organized into four 512-bit super-banks, and dual-path interconnect supporting narrow (64-bit) and wide (512-bit) data flows. Matrix A streams via eight narrow input ports, Matrix B loads through a single wide port, and the core performs fused multiply-accumulate operations with integrated SIMD quantization (scale and zero-point subtraction). The accelerator outputs int8 or int32 results via four wide ports, controlled by CSR registers for mode selection and SIMD enable. This architecture demonstrates high-throughput tensor compute with memory-efficient banking and flexible precision support, ideal for embedded AI inference and edge ML workloads. Fork and customize this diagram on Diagrams.so to explore memory bandwidth trade-offs, MAC array scaling, or quantization pipeline variations. The narrow-wide MUX selector and complex interconnect topology exemplify bandwidth-optimized designs balancing compute density with memory access patterns.