I use to have this problem. I unplug the adapter from power socket and wait on minuet. Restart the unit and the problem will go away.
You can read this post for more detail:
Please check out our recent announcement about NVFP4 performance: NVFP4 Performance Update
Good work. Looking forward to fully unlocking the DGX Spark’s performance — right now it’s only running at about half its potential.
lol 100% this!!
bumping this up again
The b12x performance on NVIDIA/Qwen3‑30B‑A3B‑NVFP4 is 69.60 tokens per second, essentially on par with Marlin, but still far from the theoretical upper limit of 182 tokens per second, leaving significant room for improvement.
I asked a similar question about the 1petaflop performance. I wanted to ask around if anyone has every observed/measured it, and if so how, but the post didn’t have any traction:
It’s unacceptable for a world-class hardware company with a complete ecosystem to engage in false advertising. This will forever brand the company with shame. Building trust takes countless hours, but destroying it can be done with just a single instance of false advertising!
More like a Mega-flop…
IMHO, Jensen is a bad cook and a lousy cartel boss. We want the blue stuff, but he keeps giving us half-baked products while raising the price. The result? People around the world are building homelabs just to make the product usable.
In some ways, Jensen has been like this since the founding of Nvidia. I still remember Nvidia’s first graphics chip and performance was not great. It was not Nvidia that kicked off the 3D gaming revolution. That honor belongs to 3Dfx, which pushed the graphics market from 2D into the 3D era.
The original 3Dfx Voodoo Graphics card was a strange product. It was initially positioned for coin-operated arcade machines, but it found far greater success with consumer gamers. You even had to pair it with a separate 2D graphics accelerator card. Despite that awkward setup, it moved the entire market forward.
3Dfx had some success in North America, but its cards were too expensive. To bring costs down, they began recruiting Taiwanese manufacturers. I worked for one of those companies. Back then, we had only ten people, so we took the quick-and-dirty route: we used the 3Dfx reference design, swapped in lower-cost parts from suppliers we could actually source, and became one of the first companies able to ship from Taiwan.
I cold-called around 50 companies in Europe and eventually found Guillemot International. At the time, Guillemot was talking to Orchid Technology about Voodoo Graphics cards, but Orchid wanted US$185 per card. I offered Guillemot US$135 and took the business away from Orchid.
But we all know what happened to 3Dfx. Jensen focused Nvidia’s energy on what the market actually wanted and put resources behind the real money-makers. He also learned to leverage other people’s resources whenever possible — just look at companies like EVGA. In the end, Nvidia caught up with 3Dfx and eliminated them.
And Jensen knows very well that companies like AMD, Google, or any other serious platform player could try to do the same thing to him. That is why Nvidia moves the way it does: stay focused, control the stack, use partners when useful, and never let a competitor get too comfortable.
That is Jensen in a nutshell: killer business sense, ruthless survival instinct, and a willingness to do the absolute minimum required to make Nvidia successful in the cutthroat computer business. He may not always serve a fully cooked product, but he knows exactly how to keep the kitchen open.
I also developed on those platforms. I recall being somewhat taken aback the amount chaos involved, as well as how far behind their platform was on stack as far as using recent versions of the standard frameworks it used (a major impediment).
I guess I should give credit that they have come a very long way from that with the DGX.
Block scaling at the hardware level for the Spark? This is music to my ears! I have 4 x DGX Sparks and I am prototyping on them for big training workloads in the cloud. Block scaling will just make iterating that much faster for me between my capacity block allocations in the cloud. So, I am guessing the Transformer Engine will be getting kernels for MXFP4 and MXFP8 for sm121 soon then?
I’ve had nothing but headaches with this device. From.the lack of software to the baby Linux distro installed.
Why does it come with a full gnome and gnome keychain but designed to be a devkit???
I’ve spent hours now trying to disable gnome and auto login the WiFi and now it’s bricked. Honestly wish I had stuck to the Mac Studio.
Use
sudo systemctl set-default multi-user.target
to disable GUI. Also, in the similar way I disabled dsx-*.service. And use
nmcli device wifi connect SSID
to connect to your WIFI. It is required once and work for many sessions but sometime reset and need to be executed via terminal again. But probably you can put it to boot sequence.
I just wish we could use the Ubuntu server edition, it’s yet more proof Nvidia had little regard for the usability of this product.
@veelacleave you can use Ubuntu Sever and customize it with DGX Software. See details at:
The criticism around NVFP4 support is fair, but some comments on this thread are overly negative. Personally I’m getting a lot of good use out of my sparks. I like them because they’re priced at a reasonable level given the market and the clustering capability is a strong USP. There are not many solutions that allow you to run 300-400b parameter models at roughly 7k of expense.
I agree – I think NVFP4 + MTP + PrismaQuant has made great strides recently.
The issue is not that the Spark is a bad device. It’s that NVIDIA has not cared enough to commit the resources needed to fully unlock NVFP4 on the device. NVFP4 was a core part of Nvidia’s value proposition - used as a main selling point due to the bandwidth cap.
The Spark is still a great device, bandwidth cap aside, but mostly thanks to fast developments in areas that benefit it and the outstanding community around it.