esp32 tx_power tuning

[mikeysklar]

Credit to "NobodySpecial" and @dhalbert for bringing this up on the discord channel circuitpython-dev.

I compared a Feather ESP32-S3 to a Qt Py ESP32-S3 running a simple network benchmark (both with PSRAM running CircuitPython 10.3.0-alpha3).

The Feather had a better signal and as tx_power went down it actually drew more current (retries?) Default tx_power of 20 seems right for this board.

The Qt Py had significantly less signal strength and drew significantly more current at tx_power 20 than tx_power 9 (+50 mA).

tx_power	Feather MB/s	QT Py MB/s	Feather mA	QT Py mA	Feather RSSI	QT Py RSSI
20	0.43	0.39	170	240	-53	-69
18	0.42	0.41	170	200	-52	-68
15	0.43	0.45	190	220	-53	-69
13	0.47	0.41	190	200	-52	-70
11	0.45	0.45	210	200	-55	-68
9	0.43	0.48	240	190	-53	-69

Now that Qt Py is already tuned for tx_power 15, but based on this one test tx_power 9 would be better.

Boards were approximately 40' away from access point which was in another room.

bench-download.py

[dhalbert]

This is very interesting. The inverse power consumption is confounding.

Could you test the C6 Feather in the same way? I was surprised NobodySpecial had trouble with the S3.

The RSSI measurements are how well the boards hear the access point. Can you get corresponding measurements about how well the access point hears them? That might not be easy. Another test would be to do RSSI measurements between our boards, one being an access point.

[mikeysklar]

I can pull the AP RSSI numbers and test with a Feather C6.

You will want to review the results of a slightly different test. The first matrix benchmarks are based on download. These are based on upload using a catcher python script on my macOS box to receive data from the S3 boards.

tx_power	Feather KiB/s	QT Py KiB/s	Feather mA	QT Py mA	Feather RSSI	QT Py RSSI
20	485	193	230	310	-46	-68
18	512	43	230	290	-47	-71
15	439	41	240	290	-48	-69
13	489	23	240	290	-47	-73
11	445	14	240	270	-48	-71
9	404	5	240	240	-46	-70

Increasing tx_power had a large effect on the Qt Py upload throughput.

The Feather is using much 30% less current that the Qt Py when they are both at their max tx_power setting.

bench-catcher.py
bench-upload.py

[mikeysklar]

My AP is lame and won't give me client RSSI values.

The Feather C6 really performs and is energy efficient on the download. Same setup location and CircuitPython build.

download:

tx_power	S3 MB/s	S3 mA	QT Py MB/s	QT Py mA	C6 MB/s	C6 mA
20	1.34	170	1.24	240	1.45	120
18	1.45	170	1.45	200	1.58	160
15	1.02	190	0.91	220	1.58	150
13	1.02	190	1.09	200	1.45	150
11	1.24	210	1.02	200	1.45	140
9	1.34	240	0.97	190	1.45	80

upload:

tx_power	S3 MB/s	S3 mA	QT Py MB/s	QT Py mA	C6 MB/s	C6 mA
20	0.47	230	0.19	310	0.21	170
18	0.50	230	0.04	290	0.22	150
15	0.43	240	0.04	290	0.19	160
13	0.48	240	0.02	290	0.14	140
11	0.43	240	0.01	270	0.12	160
9	0.39	240	0.005	240	0.04	160

[grgrant]

Sure is interesting that the S3 is so markedly better on upload throughput than the others. I wonder what could explain it. And the C6 has the same 'better' antenna that the S3 does but is less that 1/2 the throughput.

[mikeysklar]

@grgrant - The S3 is dual core vs the C6 which is single core. That can justify the larger current draw and higher throughput.

[grgrant]

Out of curiosity does your AP support WiFi6. The dual vs single core may gate the ability of the MCU to pump out data but the WiFi6 should support better over-the-air throughput.

Anyway, just curiosity. It is really nice to have numbers like yours. I've done some basic testing of different antennas and found that numbers like RSSI vary widely over time so it seems like multiple tests and averaging are good.

[mikeysklar]

@grgrant - please try the bench-download attached above or if you are feeling experimental the bench-upload with host OS catcher. It would be good to have more confirmation.

I'm not ready for the legacy incompatibilities that I'd have to fix to run IPv6. Maybe if I had a spare AP. Probably not realistic ask for users at this time.

[grgrant]

@mikeysklar - Really nice set of test code, I'm definitely going to add it to my bag of testing tricks.

Edit: I fixed the WiFi V5 note to say WiFi V4 and added IEEE protocol stds.
More interestingly I posted a second download from the Feather S3 to show wide variation on timings on runs. I'm going to setup a local http server to serve the download file to remove any Internet or foreign server interaction and will post subsequent runs.

I wish I had a C6 MCU here. It is the first Expressif to do WiFi V6 protocol (802.11 ax) vs the similarly named IPV6 but different network layer. My TP-LINK EAP-783 AP supports and does WiFi protocol V6 and if I had a C6 I could see if it is any faster than the earlier WiFi V4 (802.11 b/g/n)) protocol that the other MCUs run.

My AP is a ceiling mount in an upstairs hallway and I'm downstairs so about 16' vertically through a ceiling and down a hall.

Feather S3

Adafruit CircuitPython 10.2.1 on 2026-05-13; Adafruit Feather ESP32S3 4MB Flash 2MB PSRAM with ESP32S3
>>> import bench_download
Adafruit Feather ESP32S3 4MB Flash 2MB PSRAM with ESP32S3
ip 192.168.5.232
tx_set tx_act rssi best_KiB/s
  20  20.00   -43   1482.6
  18  18.00   -43   1617.5
  15  15.00   -43   1270.9
  13  13.00   -43   1617.4
  11  11.00   -43   1270.8
   9   9.00   -44   1270.8
SWEEP_DONE

Second bench_download
>>> import bench_download
Adafruit Feather ESP32S3 4MB Flash 2MB PSRAM with ESP32S3
ip 192.168.5.232
tx_set tx_act rssi best_KiB/s
  20  20.00   -45   1270.9
  18  18.00   -45   1270.9
  15  15.00   -45   1482.6
  13  13.00   -45   1482.6
  11  11.00   -46   1186.1
   9   9.00   -45   1617.4
SWEEP_DONE

>>> import bench_upload
Adafruit Feather ESP32S3 4MB Flash 2MB PSRAM with ESP32S3
ip 192.168.5.232
tx_set tx_act rssi up_KiB/s
  20  20.00   -43    415.6
  18  18.00   -43    373.0
  15  15.00   -43    417.2
  13  13.00   -43    425.2
  11  11.00   -43    424.4
   9   9.00   -44    420.8
UPLOAD_DONE

QtPY

soft reboot
code.py output:
Adafruit QT Py ESP32-S3 4MB Flash 2MB PSRAM with ESP32S3
ip 192.168.5.95
tx_set tx_act rssi best_KiB/s
  20  20.00   -68   1046.6
  18  18.00   -68   1186.2
  15  15.00   -68   1046.6
  13  13.00   -68   1482.6
  11  11.00   -68   1112.0
   9   9.00   -67   1482.6
SWEEP_DONE

Code done running.

soft reboot
code.py output:
Adafruit QT Py ESP32-S3 4MB Flash 2MB PSRAM with ESP32S3
ip 192.168.5.95
tx_set tx_act rssi up_KiB/s
  20  20.00   -67    284.4
  18  18.00   -69    292.0
  15  15.00   -67    237.2
  13  13.00   -71    192.5
  11  11.00   -69    225.7
   9   9.00   -70     42.4
UPLOAD_DONE

Code done running.

[bablokb]

Testing wifi performance is very difficult, since you have so much "noise" that influences the results. My wifi-router even shuts down Wifi5 if it detects flight-radar (due to official regulations). Depending on where you live and your environment, you might be testing how many other devices currently use your air-time and not the capability of your chips.

[dhalbert]

You can measure RSSI by setting up each board to test as an access point. Then write a scanner on another board that scans all the access points and reports the RSSI. Then vary the tx_power on the access point board.

The scanner could be in CircuitPython or it could just be your host computer doing access point scanning, if it report RSSI.

I am not that worried about WiFi 6, etc. I am just interested in basic RSSI measurement and whether it is correlated with tx_power.

[mikeysklar]

I used a Feather ESP32-S3 as the AP and connected with the Feather-C6 and Qt Py ESP32-S3. The boards were within 12" of each other.

Drop tx_power value and the AP sees the client with lower RSSI. That part is working as expected.

tx_power	Feather-C6 downlink	Feather-C6 uplink	QtPy-S3 downlink	QtPy-S3 uplink
20	−42	−42	−52	−53
18	−42	−42	−52	−55
15	−42	−46	−53	−56
13	−42	−48	−52	−57
11	−42	−52	−52	−58
9	−42	−56	−52	−59

test scripts

client-txpower-rssi.py
ap-txpower-rssi.py

[dhalbert]

Interesting that the RSSI decrease does not correspond to the tx_power decrease as closely as I would expect.

Looking at the throughput benchmarks, I think it would be good to repeat the measurements at a particular power multiple times to get an average, and also see how much variance there is. Depending on the wifi interference environment, there might be a lot of variance.

[mikeysklar]

Averaging 3x data upload. Using the following two boards bother going through tp-link AP and pushing data up to my macOS box. The boards are both angled so chip is facing AP.

• Qt Py ESP32-S3
• Feather ESP32-C6

The Qt Py peaked with higher throughput, but also falls apart at lower tx_power values. Using KB/s.

Qt Py ESP32-S3 averaged upload data

tx_power	RSSI	run 1	run 2	run 3	mean	min–max
20	−68	303.1	305.0	223.7	277.3	223.7–305.0
18	−69	190.4	94.6	118.7	134.6	94.6–190.4
15	−70	123.3	101.7	152.2	125.7	101.7–152.2
13	−68	84.8	90.9	93.6	89.7	84.8–93.6
11	−68	66.3	66.3	70.7	67.8	66.3–70.7
9	−68	20.6	12.5	21.1	18.1	12.5–21.1

Feather ESP32-C6 averaged upload data

tx_power	RSSI	run 1	run 2	run 3	mean	min–max
20	−55	227.1	239.6	232.0	232.9	227.1–239.6
18	−53	245.1	244.6	236.5	242.0	236.5–245.1
15	−53	190.2	188.1	194.5	190.9	188.1–194.5
13	−54	175.7	179.0	179.4	178.0	175.7–179.4
11	−53	127.3	111.1	106.7	115.0	106.7–127.3
9	−56	58.6	60.0	53.7	57.4	53.7–60.0

Side-by-side

tx_power	QtPy-S3 (RSSI ~−68)	Feather-C6 (RSSI ~−54)
20	277	233
18	135	242
15	126	191
13	90	178
11	68	115
9	18	57

avg_upload.py

[mikeysklar]

RSSI averaging with a Feather ESP32-S3 AP (board to board testing) while stepping down tx_power.

Feather ESP32-C6 as client

tx_power	downlink samples	dl mean	uplink samples	ul mean
20	−31 −31 −31 −32 −32	−31.4	−31 −31 −31 −31 −31	−31.0
18	−31 −31 −31 −31 −31	−31.0	−31 −31 −31 −31 −31	−31.0
15	−32 −32 −32 −32 −32	−32.0	−32 −34 −34 −34 −34	−33.6
13	−32 −32 −32 −32 −30	−31.6	−35 −37 −38 −38 −38	−37.2
11	−32 −32 −32 −31 −31	−31.6	−39 −41 −42 −42 −42	−41.2
9	−32 −32 −32 −31 −31	−31.6	−43 −44 −45 −46 −46	−44.8

QtPy ESP32-S3 as client

tx_power	downlink samples	dl mean	uplink samples	ul mean
20	−47 −47 −47 −48 −48	−47.4	−51 −50 −50 −50 −50	−50.2
18	−48 −48 −48 −48 −48	−48.0	−50 −50 −50 −50 −51	−50.2
15	−47 −47 −47 −47 −47	−47.0	−52 −52 −52 −52 −53	−52.2
13	−48 −47 −47 −47 −48	−47.4	−54 −55 −55 −55 −55	−54.8
11	−48 −48 −48 −48 −48	−48.0	−56 −57 −57 −58 −58	−57.2
9	−48 −48 −48 −48 −48	−48.0	−59 −59 −59 −59 −59	−59.0

side-by-side uplink only

tx_power	QtPy-S3 uplink	Feather-C6 uplink
20	−50.2	−31.0
18	−50.2	−31.0
15	−52.2	−33.6
13	−54.8	−37.2
11	−57.2	−41.2
9	−59.0	−44.8

avg_rssi.py