270 Neurocore
270 : Neurocore
- Author: Kyrylo Kalashnikov
- Description: 2x2 systolic array multiplier using 16bit floats
- GitHub repository
- Open in 3D viewer
- Clock: 20000000 Hz
How it works
This is a an implementation of a classic systolic array multiplier with uart interface.
This implementation can multiply 2x2 matrix by another 2x2 matrix. The calculations are done in float16.
How to Test
-
Initialization Sequence: Begin by sending the initialization sequence
11111110to the device. This sequence prepares the device to start receiving data for matrix multiplication. -
Sending Matrix Data:
- You will need to send the elements of the two matrices you wish to multiply. Each matrix element is a float16 number, which must be transmitted as two separate 8-bit frames (high byte first).
- Since you are multiplying 2x2 matrices, you must send a total of 8 float16 numbers, equating to 16 data frames of 8 bits each.
- For clarity, send the matrix elements in row-major order. For example, if your matrices are A and B, with elements
a11, a12, a21, a22for A and similarly for B, send them in the ordera11, a12, a21, a22, b11, b12, b21, b22.
-
Receiving the Result:
- Upon completion of the data processing, the device will first send back an acknowledgment sequence
11111110, indicating that the multiplication process is complete and the device is about to send the result. - Following this, expect to receive the result of the matrix multiplication in a similar format to the input. The device will send 4 float16 numbers (representing the resulting matrix elements) as 8-bit frames (high byte first), which you will need to interpret accordingly.
- Upon completion of the data processing, the device will first send back an acknowledgment sequence
-
Interpreting the Results:
- Collect the 8-bit frames received from the device and reconstruct them into float16 numbers to obtain the resulting matrix elements.
- These elements represent the resultant matrix from the multiplication of the two input matrices.
IO
| # | Input | Output | Bidirectional |
|---|---|---|---|
| 0 | RX input | RX output | Block multiply done status |
| 1 | Calculation start signal | ||
| 2 | Send State bit 0 | ||
| 3 | Send State bit 1 | ||
| 4 | Send State bit 2 | ||
| 5 | Send State bit 3 | ||
| 6 | Done send signal | ||
| 7 | Send data signal |