AI Accelerator RTL Design Intern
KLE TECHNOLOGICAL UNIVERSITY · TECHXCELERATE 2026 · BELAGAVI
- Designed a 4x4 weight-stationary systolic array (signed INT8 inputs, 32-bit accumulation) in SystemVerilog with parameterized modules for the PE, PE array, weight loader, activation skew, output collector, and control FSM.
- Pipelined multiply-accumulate datapath with diagonal activation skew and FSM-controlled preload, compute, and drain phases; one 4x4 multiply completes in 18 clock cycles (1.5 µs at 12 MHz).
- Implemented on iCE40UP5K (VSDSquadron FPGA Mini) with Yosys and nextpnr at 12 MHz, with a UART interface, power-on reset, and cycle counter; 16 multipliers mapped to LUTs since the array exceeds the 8 DSP blocks.