OpenAI reports that a Jalapeño matrix-unit multiplier used about 15% less synthesized cell area and fell from two clock cycles to one. A separate adder saved roughly 15% area while remaining at two cycles. These are circuit changes, not whole-chip gains; elapsed time also depends on clock period. Aggregate estimated standard-cell savings were 13% in the vector unit and 10% in the matrix unit.
The AMD work reports up to 14% improvement in a combined power-performance-area measure against heavily optimized GPU blocks. Without its weighting, that cannot be translated into 14% more speed.
The RISC-V experiment allows area up to sixteen times the original core. Failed simulations score zero; passing candidates combine hidden-workload IPC with synthesis-estimated frequency. Starting from specifications scored higher than editing RTL, but could omit original features. Preserving required features and using tighter area limits and physical-synthesis timing would make the comparison more representative of production.