Memory Map Reference

Address Space

0x00000000  ┌─────────────────────┐
            │   ITCM              │  8 KB (0x2000 bytes)
            │   .text             │  Machine code
            │   .rodata           │  Read-only data (constants)
            │   .init_array       │  C++ static constructors
0x00002000  ├─────────────────────┤
            │   (unmapped gap)    │
0x00010000  ├─────────────────────┤
            │   DTCM              │  32 KB (0x8000 bytes)
            │   .data             │  Initialized global variables
            │   .bss              │  Zero-initialized globals
            │   .heap             │  Dynamic allocation (grows up)
            │   .stack            │  Stack (grows down from top)
0x00018000  ├─────────────────────┤
            │   (unmapped gap)    │
0x20000000  ├─────────────────────┤
            │   EXTMEM            │  4 MB (0x400000 bytes)
            │   .extdata          │  Overflow initialized data
            │   .extbss           │  Overflow zero-initialized data
0x20400000  └─────────────────────┘

Memory Regions

Region Base Size End Contents

ITCM

0x00000000

8 KB

0x00002000

Code, read-only data, init arrays

DTCM

0x00010000

32 KB

0x00018000

Data, BSS, heap (up), stack (down)

EXTMEM

0x20000000

4 MB

0x20400000

Overflow data (higher latency)

Linker Script Sections

From coralnpu_tcm.ld.tpl:

ITCM Sections (Code Memory)

Section Contents Alignment

.text

Machine code (main(), loops, CRT startup)

4 bytes

.rodata

Read-only constants (weight arrays, string literals)

4 bytes

.init_array

C++ static constructor pointers

4 bytes

.fini_array

C++ static destructor pointers

4 bytes

DTCM Sections (Data Memory)

Section Contents Alignment

.data

Initialized global variables (input/output arrays with __attribute__((section(".data"))))

4 bytes

.sdata

Small initialized data (accessed via global pointer)

4 bytes

.bss

Zero-initialized globals (cleared by CRT startup)

4 bytes

.sbss

Small zero-initialized data

4 bytes

.heap

Dynamic allocation area (grows upward from _heap_start)

8 bytes

.stack

Stack area (grows downward from _stack_top)

16 bytes

EXTMEM Sections (External Memory)

Section Contents Alignment

.extdata

Large initialized data that overflows DTCM

4 bytes

.extbss

Large zero-initialized data that overflows DTCM

4 bytes

Linker Symbols

The linker script exports symbols used by CRT startup and the simulator:

Symbol Type Purpose

_text_start

Address

Start of .text section

_text_end

Address

End of .text section

_data_start

Address

Start of .data section in DTCM

_bss_start

Address

Start of .bss (cleared by CRT)

_bss_end

Address

End of .bss

_heap_start

Address

Bottom of heap (for _sbrk())

_stack_top

Address

Top of stack (initial SP value)

_ret

Variable

Return value from main() (written by CRT)

__global_pointer$

Address

Global pointer for gp register

input_0, input_1

Variable

Input arrays (accessed by simulator)

output_0, output_1

Variable

Output arrays (read by simulator)

Memory Budget Example: RGB-to-Grayscale

Item Region Size (bytes) Size (floats)

input_0 (1×3×4×4)

DTCM

192

48

output_0 (1×1×4×4)

DTCM

64

16

v0 (weights)

ITCM

12

3

Code (main + CRT)

ITCM

~2,000

—

Stack

DTCM

~1,024

—

Total ITCM

~2,012 / 8,192 (25%)

Total DTCM

~1,280 / 32,768 (4%)

Memory Budget Example: 224×224 RGB Image

Item Region Size (bytes) Fits?

Input (1×3×224×224)

DTCM

602,112

No (32 KB limit) — must use EXTMEM or tiling

Output (1×1×224×224)

DTCM

200,704

No — must use EXTMEM or tiling

Weights (1×3×1×1)

ITCM

12

Yes

For production models, tiling is required: process the image in spatial tiles (e.g., 16×16) that fit in DTCM, writing results back to EXTMEM between tiles.

I/O Convention

The NPU uses a shared-memory I/O model:

  1. Simulator writes input data to the input_0 symbol address in DTCM

  2. Program executes main(), reading from input_0 and writing to output_0

  3. Simulator reads output data from the output_0 symbol address

  4. No syscalls, no file I/O, no interrupts — pure memory-mapped data exchange