[REPORT] Cubie A7A: RGMII CRC errors resolved by changing drive level from 1 to 2
From: Sergey Kacheev
Date: Fri Nov 06 2026 - 16:09:16 EST
Hi,
I investigated Gigabit Ethernet CRC errors on a Radxa Cubie A7A
(Allwinner A733, Maxio MAE0621A PHY).
On this board, changing the RGMII pin drive setting from hardware
Level 1 to Level 2 eliminated CRC errors in repeated tests.
Environment:
- Kernel: 6.6.98-5-aw2511-srv, based on the vendor BSP
- MAC driver: vendor dwmac-sunxi / stmmac
- PHY ID: 0x7b744412
- Local backport of the proposed Maxio PHY driver, with the vendor
EEE-disable workaround
- Link: 1000 Mbps, full duplex, autonegotiation enabled, EEE disabled
This is a vendor-kernel result; I have not reproduced it on mainline.
The relevant pin group covers PH0-PH15. The DT change was:
allwinner,drive = <1>;
->
allwinner,drive = <2>;
The BSP pinctrl driver labels these settings as 20 mA and 30 mA.
However, I am deliberately reporting hardware drive levels rather
than treating those labels as verified electrical current ratings.
The A733 User Manual V1.00 documents four levels for PH_DRV0/PH_DRV1
(sections 18.14.6.99 and 18.14.6.100). Section 18.14.3.4 describes
Level 1 and Level 2 as enabling two and three parallel output
branches, respectively, reducing the nominal output resistance
from r0/2 to r0/3.
With TX delay = 8 and RX delay = 10 held constant, repeated tests
showed:
- Level 1: CRC errors with pseudorandom Ethernet payloads and
ordinary TCP transfers.
- Level 2: no CRC errors in the same tests.
- Returning to Level 1: CRC errors returned.
For example, a 300,000-frame pseudorandom test produced 141 CRC
errors at Level 1, zero at Level 2, and 61 after returning to
Level 1.
Payload choice mattered: fixed 00/ff/aa/55 patterns did not expose
the problem in earlier tests, while pseudorandom payloads did.
Validation at Level 2 included:
- 4 GiB of bidirectional TCP transfers before reboot and another
4 GiB after reboot.
- 600,000 raw frames per validation run, covering frame lengths
of 64, 128, 256, 512, 1024 and 1514 bytes.
- No CRC counter increments on either the Cubie MAC or the
connected switch during these runs.
- TCP throughput around 111-112 MB/s.
A separate TX-delay 8 -> 9 -> 8 comparison at Level 2 was also
CRC-free. The small throughput difference was not sufficient to
establish a meaningful advantage for either value.
The persistent configuration currently uses drive Level 2,
TX delay = 8 and RX delay = 10.
These results are from one board. I have no oscilloscope
measurements, so they establish the observed effect of the drive
setting, not the precise electrical cause or a board-wide fix.
Would Level 2 be an appropriate board-specific RGMII pinctrl
default for the Cubie A7A? Also, is there an authoritative mapping
between the A733 drive levels and their electrical characteristics?
The BSP's 10/20/30/40 mA labels do not appear to match the GPIO
current figures in the datasheet.
References:
A733 User Manual V1.00:
https://gitlab.com/tina5.0_aiot/product/docs/-/blob/product-aiot-stable/A733/Hardware硬件类文档/芯片手册/A733_User%20Manual_V1.00.pdf
A733 Datasheet V1.01:
https://gitlab.com/tina5.0_aiot/product/docs/-/blob/product-aiot-stable/A733/Hardware硬件类文档/芯片手册/A733_Datasheet_V1.01.pdf
BSP pinctrl driver:
https://github.com/radxa/allwinner-bsp/blob/f9a496c24bf6b7cc31c4b5c716ece58b2c287f01/drivers/pinctrl/pinctrl-sunxi.c
Best regards,
Sergey Kacheev