Deployment Hardware Matrix
S3-WROOM-1
Model Backbone
~250K params
Input Tensor Resolution
9,216 px
Quantization Format
~300KB Flash
Kernel Acceleration
Xtensa S3 SIMD
WiFi / MQTT Radio Stack
+0KB SRAM
Power & Trigger Strategy
Internal 512KB SRAM Memory Pool
Safe (Fits internal SRAM)
Tensor Arena (140 KB)
TFLM Runtime (22 KB)
App Logic (30 KB)
Camera DMA & Core (100 KB)
WiFi/MQTT (0 KB)
Free SRAM (220 KB)
* Note: Keeping the tensor arena inside internal SRAM ensures full 240MHz zero-wait-state bus bandwidth. Placing it in external PSRAM collapses FPS to ~3 FPS. Frame buffer (QVGA 320x240 RGB565 = ~150KB) safely resides in 8MB PSRAM.
Pipeline Latency Waterfall (Person Detection)
Local is 5.1x Faster
OV2640 DMA 25ms
Downsample 15ms
ESP-DL Invoke 53ms
GPIO 2ms
PSRAM Overwritten in 100ms
Zero Hallway Frames Leave Premise
JPEG 75ms
WiFi TX 135ms
Cloud Queue 185ms
Push ACK 85ms
Microwave interference spikes to 2,500ms
Uncompressed raw imagery stored on remote servers
Visual Wake Words Optical Horizon
At 3.0m: Person Resolves Clearly
36-Month Cost of Ownership (BOM vs Cloud API)
| Expense Item |
ESP32-S3 Local Edge |
Cloud Camera Stream |
| Hardware BOM (Camera + DevKit) |
$12.00 (One-time) |
$49.99 (Proprietary) |
| Vision Subscription (36 mo @ $9.99) |
$0.00 |
$359.64 |
| Bandwidth Egress & Storage |
$0.00 (0 MB/mo) |
$0.00 (Included) |
| Net 36-Month Cost |
$12.00 |
$409.63 |
Generated ESP-IDF / Arduino C++ Header