DGX Spark unstable internet when connected to enterprise WIFI

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Hi,

I tried to connect a NVIDIA DGX Spark to the enterprise WIFI at MIT: “MIT Secure” / “MIT” / “eduroam”.
But the internet on the DGX Spark is quite unstable.
Here are a few of my observations.

  1. I do “ping 8.8.8.8” on the DGX Spark, and see something like this:
    64 bytes from 8.8.8.8: icmp_seq=1 ttl=114 time=20.622 ms
    64 bytes from 8.8.8.8: icmp_seq=2 ttl=114 time=17.708 ms
    64 bytes from 8.8.8.8: icmp_seq=3 ttl=114 time=17.869 ms
    64 bytes from 8.8.8.8: icmp_seq=31 ttl=114 time=18.196 ms
    Note that it got packet loss after every few responses. This happens
    regardless of whether it’s on"MIT Secure" / “MIT” / “eduroam”.
    And SSH from laptop to DGX Spark would be very unstable/

  2. If I do “ping 8.8.8.8” in my laptop on “MIT Secure” / “MIT” / “eduroam”,
    There’s no packet loss.

  3. When the DGX Spark is connected to non-secure wifi like “MIT Guest“ / phone hotspot / my home WIFI, the internet is stable, and there is no Packet loss.

Is there any fix to this?

Thanks

2

Could you check wifi.powersave in NetworkManager settings

/etc/NetworkManager/conf.d/default-wifi-powersave-on.conf , set wifi.powersave = 2 to disable it?

Could you also please run: sudo journalctl -xe | grep NetworkManagerto check for errors.

3

Hello, @raphael.amorim

Per your previous instructions, I have disabled WiFi power management on the DGX Spark.

Actions Taken:

  1. Modified /etc/NetworkManager/conf.d/default-wifi-powersave-on.conf and set wifi.powersave = 2.

  2. Verified the change was read by NetworkManager in the logs.

  3. Deleted “MIT GUEST” to force a connection to the secure enterprise network “MIT”.

Current Status: The packet loss persists. The connection drops periodically during ping 8.8.8.8 tests.

Attached Log: I have attached a log (nvidia_wifi_report.txt) of past 20 minutes.

nvidia_wifi_report.txt (14.3 KB)

Please advise on the next steps. Is this a known driver incompatibility with Enterprise WPA2 networks on this hardware?

4

Let’s increase the debug logs: sudo nmcli general logging level DEBUG domains all

Could you try to connect using the command line

nmcli connection add type wifi ifname wlP9s9 con-name "MIT" \
  ssid "MIT" \
  802-11-wireless-security.key-mgmt wpa-eap \
  802-1x.eap peap \
  802-1x.phase2-auth mschapv2 \
  802-1x.identity "identity@example.com" \
  802-1x.password "YourPasswordHere"

nmcli connection up "MIT" ifname wlP9s9

Revert the log level to INFO: sudo nmcli general logging level INFO

5

Hi, @raphael.amorim
The ‘MIT’ SSID I am using is the PSK (Pre-Shared Key) version, It does not take a username.

I have successfully connected using key-mgmt wpa-psk, but the packet loss issue persists.

Attached is the DEBUG log you requested."

nvidia_debug_log.txt (1.9 MB)

6

If you’re using PSK and not an 802.1x/enterprise SSID, then we can rule out a WPA2-Enterprise issue, but I’ll verify that at the office later this week.

I was checking those two bugs, which are similar:

In one of them one person mentions Kernel 6.16 doesn’t have the issue. The spark is on 6.14 right now.
Which version are you running? uname -a

7

Linux spark-4551 6.14.0-1013-nvidia #13-Ubuntu SMP PREEMPT_DYNAMIC Wed Oct 29 06:01:19 UTC 2025 aarch64 aarch64 aarch64 GNU/Linux

8

We could check if it’s not some weird MTU/fragmentation:

nmcli connection modify "MIT" 802-3-ethernet.mtu 1400

You can also open a separated terminal and type:

sudo dmesg -w | tee /tmp/dgx-wifi-dmesg.log

And in another one:

start pings to 8.8.8.8 and your gateway
wait until you see a burst of packet loss
stop dmesg -w

Then attach /tmp/dgx-wifi-dmesg.log. This can catch some subtle MT7925 driver/firmware messages that don’t show up in NetworkManager’s own logs.

I’m also seeing some DNS issues with ipv6. Could you try to disable ipv6 on your connection:

nmcli con modify "MIT" ipv6.method ignore
nmcli con down "MIT" && nmcli con up "MIT"