1261-EX Accelerating Matrix Multiplication on a 16-Bit CPU With a Hardware Multiply-Accumulate Unit
Matrix multiplication is central to many neural-network workloads, and specialized processors accelerate it with multiply-accumulate (MAC) hardware. I am studying this idea on the K86, a 16-bit CPU designed by Kennesaw State University students in NI Multisim. Although the K86’s instruction set and assembler recognize MUL, the original ALU subcircuits do not contain an implemented multiplier. My question is how much a small arithmetic extension can reduce matrix-multiplication cycles and what additional hardware it requires. I plan to add an unsigned 8-bit multiplier, a dedicated 16-bit saturating accumulator, and two instructions. MAC multiplies the low eight bits of two registers and accumulates the product; MACR reads the previous accumulator value into a register and clears the accumulator. Configuration B’s existing MUL instruction will use the same eight-by-eight-bit multiplier and return a 16-bit product, with its limited operand width documented. The opcode spreadsheet and assembler identify 0x22 and 0x23 as unused; I still need to confirm their availability in the Multisim decoder. I will use a corrected version of the original assembler. I will compare software shift-and-add multiplication, hardware MUL followed by ADD, and fused MAC on identical matrix inputs. I will check every output against Python, measure complete-program cycles with a 32-bit simulated counter, and record the additional logic and registers relative to the original CPU. The experiment will distinguish the gain from hardware multiplication from the additional gain from fusion.
Authors
- Shivam Bansal
Institutions
- Kennesaw State University (US)
Publication Details
- Journal
- DigitalCommons - Kennesaw State University (Kennesaw State University)
- Published
- 2026-11-20
- Primary Topic
- Numerical Methods and Algorithms
- Type
- article
- Field-Weighted Citation Impact
- 0.00