MoonMath challenge
Matrix multiplier machine
Design a machine that multiplies two 1024 × 1024 FP16 matrices through an external-memory interface.
The machine has access to external memory containing two matrices of 16-bit elements. It can read four elements (64 bits) and write four elements (64 bits) every clock cycle.
Matrix A is located at address 0 and above. Matrix B is located at address 256K and above. Write the result to address 512K and above.
Interface
- Input: Clock (1b)
- Input: Reset (1b)
- Output: mem_read_en (1b)
- Output: mem_read_addr (20b)
- Input: mem_read_data (64b), arriving 3 clocks after mem_read_en is set
- Output: mem_write_en (1b)
- Output: mem_write_addr (20b)
- Output: mem_write_data (64b)
- Input: multiply_en (1b)
- Output: ready (1b)
When multiply_en is asserted
- Deassert ready.
- Multiply the two matrices and write the result to address 512K and above.
- Assert ready.
Design requirements
- Build a machine that executes the task as fast as possible; area is not a concern.
- Build a separate machine that executes the task using only four multipliers.