Just recived today a new DGX Spark. First boot, complet update, wifi working but when I try to use the ethernet connection it stalls every time making almost impossible to download any model. I have disabled eee, changed cable and switch with no luck. Some advice?
Hi @yoprogramo ,
Thanks for the detailed report, and for already ruling out EEE, the cable, and the switch.
This does not appear to be a currently documented DGX Spark known issue, so I’d treat it as a link negotiation / driver / image-state problem rather than expected behavior.
A few things I’d try next:
-
Confirm the system is fully updated, then reboot once more. The current DGX Spark Founders Edition software stack is DGX OS 7.4.0, driver 580.142, CUDA 13.0.2, kernel 6.17, and UEFI 1.107.26.
-
Test with Wi-Fi temporarily disabled, just to make sure traffic is not bouncing between interfaces while you’re downloading models.
-
Please capture the following and post the output:
ip -br link
ip addr
ethtool -i <ethernet_iface>
sudo ethtool <ethernet_iface>
ip -s link show <ethernet_iface>
sudo dmesg -Tw | grep -Ei 'eth|r8169|realtek|link'
- As a quick isolation step, force the port to 1Gb/full duplex temporarily. If that becomes stable, it points to a 10Gb negotiation/interoperability issue upstream:
sudo ethtool -s <ethernet_iface> speed 1000 duplex full autoneg off
Then test again, and afterward restore normal autonegotiation:
sudo ethtool -s <ethernet_iface> autoneg on
- If you originally completed first-time setup over Wi-Fi, power the unit off, connect Ethernet first, then boot and retest. The setup guide specifically recommends plugging in wired networking before installation to avoid connection issues later.
Also please confirm you’re using the supplied power adapter, since NVIDIA calls that out for proper system behavior and performance.
The data you asked for:
betandeal@spark-e6ca:~$ ip -br link
lo UNKNOWN 00:00:00:00:00:00 <LOOPBACK,UP,LOWER_UP>
enP7s7 UP 4c:bb:47:7d:e6:ca <BROADCAST,MULTICAST,UP,LOWER_UP>
wlP9s9 UP f8:3d:c6:6d:0c:ea <BROADCAST,MULTICAST,UP,LOWER_UP>
docker0 DOWN 0e:a8:f6:be:cf:6d <NO-CARRIER,BROADCAST,MULTICAST,UP>
betandeal@spark-e6ca:~$ ethtool -i enP7s7
driver: r8127
version: 11.014.00-NAPI
firmware-version:
expansion-rom-version:
bus-info: 0007:01:00.0
supports-statistics: yes
supports-test: no
supports-eeprom-access: no
supports-register-dump: yes
supports-priv-flags: no
betandeal@spark-e6ca:~$ sudo ethtool enP7s7
Settings for enP7s7:
Supported ports: [ TP ]
Supported link modes: 10baseT/Half 10baseT/Full
100baseT/Half 100baseT/Full
1000baseT/Full
10000baseT/Full
2500baseT/Full
5000baseT/Full
Supported pause frame use: Symmetric Receive-only
Supports auto-negotiation: Yes
Supported FEC modes: Not reported
Advertised link modes: 10baseT/Half 10baseT/Full
100baseT/Half 100baseT/Full
1000baseT/Full
10000baseT/Full
2500baseT/Full
5000baseT/Full
Advertised pause frame use: Symmetric Receive-only
Advertised auto-negotiation: Yes
Advertised FEC modes: Not reported
Link partner advertised link modes: 10baseT/Half 10baseT/Full
100baseT/Half 100baseT/Full
1000baseT/Full
Link partner advertised pause frame use: Symmetric Receive-only
Link partner advertised auto-negotiation: Yes
Link partner advertised FEC modes: Not reported
Speed: 1000Mb/s
Duplex: Full
Auto-negotiation: on
Port: Twisted Pair
PHYAD: 0
Transceiver: internal
MDI-X: on
Supports Wake-on: pumbg
Wake-on: g
Current message level: 0x00000033 (51)
drv probe ifdown ifup
Link detected: yes
I cannot disable autonegotiation:
sudo ethtool -s enP7s7 speed 1000 duplex full autoneg off
netlink error: link settings update failed
netlink error: Invalid argument
Instead I keen autonegotiating on but forcing the speed:
sudo ethtool -s enP7s7 speed 1000 duplex full autoneg on
no luck
If I force the advertisement
sudo ethtool -s enP7s7 advertise 0x020
Seems more stable, let’s try something more…
Not stable yet. The connection closes just few minutes from stablished.
To follow your advice to install with ethernet from the start. How do I reinstall from scratch?
i have a massive problem with this and NCCL and RoCE, it randomly changes the ip adresses so i have to chut down IPV6 for it too run. and a lot of MTU modificatiosn and more in NCCL for proper 160Gbit/s for one cable. Put tha in foro RJ45 and WIFI and you have a 4 channel ip foro QSFP and 2 more for ethernet and wifi.
BAD BOY NVIDIA!!! you dont do that!!!
Listen, every time you reboot your system your ip changes if you have 2 nodes or more if you are using RoCE NCCL for communication no TCP IP, you have to be very specific on how you configure and manage your bandwidth and timing.
I recommend sign GROK for technical problems, he can fix it good.
@yoprogramo if your Spark is the Founder Edition, the recovery steps are at System Recovery — DGX Spark User Guide
The latest recovery image is https://developer.nvidia.com/downloads/dgx-spark/dgx-spark-recovery-image-1.135.29.tar.gz
Finally all works as expected just disabling wifi and booting again… Just very strange.