102 Scalable Banked L1 Memory Fabric for Edge AI
102 : Scalable Banked L1 Memory Fabric for Edge AI
- Author: Vishnukumar Varatharaja Perumal
- Description: DFF-based silicon demonstrator of a scalable banked L1 weight memory fabric with DMA streaming and GEMM MAC, ported to IHP sg13g2
- GitHub repository
- Open in 3D viewer
- Open in Tiny Tapeout Explorer
- Clock: 40000000 Hz
How it works
This project is a silicon demonstrator of a scalable banked L1 weight memory fabric for edge AI accelerators, developed as part of an M.S. thesis at San Diego State University (SDSU IoT Laboratory).
The design contains four independent 4-word x 32-bit weight banks (DFF-based), a priority arbiter that grants DMA traffic priority over CPU access, a DMA streaming engine, and a weight-stationary GEMM MAC unit. The same arbiter, DMA, and GEMM RTL was verified at 33-bank scale with sky130 SRAM macros through a complete Cadence RTL-to-GDS flow (Xcelium, Genus, Innovus, Magic DRC = 0).
Writes assemble 32-bit words one byte at a time through the bidirectional bus; a write commits when byte 3 is written. Reads are combinational: select mode=read with a bank/word/byte address and the selected byte appears on the bidirectional bus.
How to test
All operations are driven through ui_in:
| Bits | Field | Meaning |
|---|---|---|
| 7:6 | mode | 00=write 01=read 10=DMA 11=GEMM |
| 5:4 | bank | bank select 0-3 |
| 3:2 | byte | byte select within 32-bit word |
| 1:0 | word | word address within bank |
Write a word: mode=00, drive each byte on uio (byte_sel 0,1,2,3). The write commits on byte 3. Read it back with mode=01: the selected byte appears on uio_out.
In DMA mode (10), byte_sel selects the DMA register: 0=source address, 1=transfer length, 2=start. The register is written with the value currently in the 32-bit write accumulator (fill it first with a mode-00 write sequence). A transfer streams bank-0 words into the GEMM unit; dma_grant (uo_out[5]) asserts during the transfer and gemm_done_irq (uo_out[7]) fires when the MAC completes.
In GEMM mode (11), byte_sel=0 presents the accumulated MAC result: the low byte on uio_out and its low nibble on uo_out[3:0]. Example: weight 0x00030002 streamed 4 times gives 4x(2x3) = 0x18.
External hardware
None required. All testing is done through the Tiny Tapeout demo board I/O.
IO
| # | Input | Output | Bidirectional |
|---|---|---|---|
| 0 | mode[1] / bank_sel[1] / byte_sel[1] / word_sel[1] | gemm_result[0] | data[0] |
| 1 | mode[0] / bank_sel[0] / byte_sel[0] / word_sel[0] | gemm_result[1] | data[1] |
| 2 | bank_sel[1] | gemm_result[2] | data[2] |
| 3 | bank_sel[0] | gemm_result[3] | data[3] |
| 4 | byte_sel[1] | cpu_grant | data[4] |
| 5 | byte_sel[0] | dma_grant | data[5] |
| 6 | word_sel[1] | dma_done_irq | data[6] |
| 7 | word_sel[0] | gemm_done_irq | data[7] |