RTL8721Dx vs ESP32-S3 CSI comparison — RTL8721Dx has the only MEASURED throughput number

[ruvnet]

Summary

Detailed comparative analysis of RuView's two real (non-simulated) hardware WiFi CSI implementations: the existing ESP32-S3 CSI node and the new RTL8721Dx (Realtek AmebaDplus) CSI node added this session (ADR-323, branch feat/rtl8721dx-csi-bringup, commit 72bfb5059).

Full report: https://gist.github.com/ruvnet/554751b1068c48f7ee210aa447cd4088

Headline findings

1. Both are 1×1 SISO — neither has spatial diversity from a single chip; that requires multiple physical boards on either platform.
2. RTL8721Dx has a genuinely passive CSI mode (RTW_CSI_MODE_NORMAL, samples real AP beacons/data with zero injected traffic). ESP32's CSI capture always self-injects probe/NDP frames to guarantee reports (csi_collector.c:383-384) — there's no pure-passive equivalent.
3. The starkest asymmetry: measured performance. RTL8721Dx now has a MEASURED, repeatable real-hardware benchmark (scripts/benchmark-rtl8721dx-csi.py): 14.67 fps baseline → 66.40 fps tuned. ESP32 CSI has never been benchmarked on real hardware in this repo — the "~20 Hz" / "20-100 Hz" figures in its docs are configuration values, not measurements, and ADR-028's own audit flags a separate "54,000 fps" claim elsewhere as unverified.
4. ESP32 is far more mature as a product: OTA, a 4-tier on-device DSP pipeline (up to WASM-based programmable sensing), TDM multi-node scheduling (ADR-029 — which directly solves a channel-contention problem the RTL8721Dx work hit and left unsolved this session), QEMU CI with fuzzing. RTL8721Dx is currently a raw CSI streamer with none of that.
5. A real, vendor-independent bug was found in the process: wifi-densepose-sensing-server --source auto silently never binds the UDP CSI receiver on any host machine that has its own WiFi adapter — affects ESP32 ingestion equally, not just Realtek.

Suggested follow-up

The most actionable next step: build an ESP32 equivalent of scripts/benchmark-rtl8721dx-csi.py and run it against real ESP32-S3 hardware, so the throughput comparison above stops being "one side was measured, one wasn't" and becomes a real apples-to-apples number.

Co-Authored-By: claude-flow ruv@ruv.net

[ruvnet]

The ESP32 counterpart of benchmark-rtl8721dx-csi.py suggested above is in #2050 as scripts/benchmark-esp32-csi.py. It parses ADR-018 frames and reports fps, sequence loss (with per-gap attribution, so on-device drops can be told apart from over-the-air loss), and every CSI outage with a timestamp.

First MEASURED ESP32 numbers (C6, one AP on channel 9, sensing traffic from self-ping plus ambient):

• stock 0.8.12: about 13–22 fps delivered, with 21–28% of sequenced frames discarded on-device by a double rate gate (fixed in fix(firmware): recover stalled CSI capture and stop discarding accepted frames (#1941) #2050)
• after fix(firmware): recover stalled CSI capture and stop discarding accepted frames (#1941) #2050: 14.8–21.5 fps delivered with 0.08–0.15% loss over 180–600 s windows

The frame rate depends on ambient traffic, so an apples-to-apples comparison with the RTL8721Dx's 66 fps figure needs both boards on the same AP at the same time. That run hasn't been done yet. No ESP32-S3 measurement yet either.