AMD Ryzen AI 300 not fully supported by Linux yet?

Framework Laptop 13 Linux
21

mt7925 also fine here for the most part on NixOS 6.15, although I did hit an intermittent probing issue, which seems to be fixed with the latest firmware:

22

Do you mean 6.15.1? I don’t see 6.15.0 in the nix packages nor in the kernel mirrors.

I wonder if the router comes into play. I’m on a BPI-R3 running OpenWrt.

Perhaps… I’m on Mikrotik hAP ac2 (running RouterOS 6.49.18).

23

I’m currently running 6.15, although 6.15.1 has just reached nixos-unstable, so will be updating.

24

I investigated the issue a bit further and found that for my setup, the regression is introduced in 6.14.3 (6.14.2 is the last version that works for me).

Also an interesting detail that I didn’t notice because I didn’t care about upload speed initially: on 6.14.3 I get around 16Mbps download and 300Mbps upload, and on 6.14.2 I get around 250Mbps download and 50Mbps upload. So on one version the download is bad but the upload is good, and vice-versa on the other version.

Reported my issue/findings here.

2 Likes
25

6.15.0

26

Just updated, now on 6.15.1, Wi-Fi speeds are the same as on 6.15.0 FWIW.

27

Between 6.14.2 and 6.14.3 several commits landed for the MediaTek 7925 which adjust things like powersave, MLO setup, etc (changelog here, search for ‘mt7925’ and ‘mt76’).

As @serafean pointed out, it’s good to check firmware versions as well. Within the past two weeks, new WiFi (and Bluetooth) firmware was recently released (late May build date, link here). Possibly that may help if the performance is similar amongst all kernels post 6.14.2.

28

If AMD or mediatek (or anyone) releases buggy drivers for windows, framework can’t do anything about it either. Case in point : recent nvidia drivers. None of the computer vendors, or game developers can do anything about that.
My view is that with Linux I at least have a chance to debug the issue, so I take that.

It can’t work: those are no longer PCIe devices, but some intel cheapskate called CNVio(2) where only the radio is on the expansion card, and all the rest of the “wifi hardware” is prebaked on the chipset.
TLDR: don’t even try intel cards ending 1.

2 Likes
29

I don’t have the Framework Laptop 13", AMD AI 300 Series, so take this with a grain of salt. I actually saw improvement with upstream stable 6.14.3 kernel compared to upstream stable 6.14.2.

Testing info below, but generally, performance seemed to improve with successive kernel versions (yay, stabilizing!). 6.15.2 with May firmware performed the best, so hopefully this improvement translates to the Framework platforms as well :slight_smile:

My testing scenario includes the following:

  • mt7925 radio in an Intel-based CPU system
  • mt7996-based WiFi 7 AP (no MLO)
  • AP configured to 40MHz at channel 36 (surprisingly unused where I’m at)
  • Station configured to WPA2, no PMF (802.11w)
  • Kernels built from tagged versions in the Linux stable tree
  • Single-direction, 1Gbps-configured iperf3 UDP traffic with 10 parallel streams (iperf3 --bind-dev wlan0 -b 1G -P 10 -u, add -R to do opposite direction)

I tested each of the following kernels with mt7925 firmware from March, April, and May (available in the Linux firmware repo on kernel.org). I skipped over a few 6.14 kernels, as there really aren’t many ‘mt76’ or ‘mt7925’ driver fixes/updates until 6.14.8 and 6.14.9 and none after.

  • 6.14.2
  • 6.14.3 (includes successive connection failing fix)
  • 6.14.9 (includes IPv6 and multicast traffic drop fix)
  • 6.15.2

For folks looking to reproduce or especially detailed info:

  • Query driver and firmware information with ethtool -i wlan0 command (use -S instead for interface statistics)
  • Link rate and related information (MCS, NSS) can be queried using the iw wlan0 link command. Note that you’ll need to get a recent version to display EHT (WiFi 7) information. I like running it in a loop like watch -n .1 "iw wlan0 link"
  • wpa_supplicant config below (wpa_supplicant -c wlan0.conf -i wlan0)
  • Using a secondary sniffer radio, saw that 6.14.2 would not do upload traffic with block ACKs (frame aggregation) with all firmware versions. 6.14.3 did, though, which seemingly explains the substantial increase in upload throughput.
# Only used if you're also using 'wpa_cli'
ctrl_interface=/var/run/wpa_supplicant

network={
    # Need to use 'wpa_passphrase' to generate WPA2 'psk'
    # e.g. 'wpa_passphrase ssid password'
    ssid="ssid"
    psk=XXXXXXX

    # Can force to specific band using BSSID
    #bssid=xx:xx:xx:xx:xx:xx

    pairwise=CCMP
    key_mgmt=WPA-PSK
    ieee80211w=0
}
30

Yup, I have an older 13" Ryzen I purchased last year, no issues with it. Only the new AI 300.

31

Yeah the Intel AX210 NGW I got has WiFi 6E, which is fine for my needs. One thing I love about these laptops is I could upgrade down the line since their not soldered to the mobo.

32

My complaint is saying “official support” for Fedora, but having barely functional WiFi. The laptop also crashed when plugged into my thunderbolt adapter.

I think there just needs to be a little more clarity around the fact this is still fairly bleeding edge ATM.

I must say with the WiFi chip swapped, the system is super stable (still issues with thunderbolt though).

33

“official support” mainly means that if you contact support, or file a ticket, with Fedora frame.work will take a look at it. With Gentoo they’ll tell me to retest with Fedora.

2 Likes
34

You also won’t be told your device is out of warranty because you installed linux. Anyone who has used linux for any length of time knows months 1-6 after a new cpu is released is the time to buy it only if you 1) absolutely have to replace a device, 2) don’t mind paper cuts, and/or 3) expect to troubleshoot and generally like doing that in order to help clean it up for everyone else.

If you don’t have time for any of that you generally get last years model.

1 Like