Repeat after me .. “DGX Spark is not a inference performance play”, “DGX Spark is not a inference performance play”, “DGX Spark is not a inference performance play”.
Thinking about getting a 2nd GB10 unit but one of the OEM versions such as Gigabyte, MSI, or Dell. I heard that connecting two GB10s from different vendors should not be a problem. However, I’ve also heard that some OEM versions only have 119GB of available RAM rather than121GB (possibly due to a missing firmware update from an OEM). Is this a concern?
Got the second and didn’t regret it. Now the jump to 4 has me tempted.
Repeat after me: “The market dictates the use case, not the marketing slide.”
While the DGX Spark wasn’t engineered as a primary inference powerhouse, the community isn’t acting out of ignorance. Users are leveraging its 128GB of unified memory to run large models that would otherwise require a prohibitively expensive array of RTX Pro 5000 or 6000 cards.
Despite the modest 270 GB/s bandwidth, recent breakthroughs in quantization and Multi-Token Prediction (MTP) are performing miracles, making high-parameter inference viable for those who prioritize capacity over raw speed.
It’s the same story with the RTX 5090. NVIDIA didn’t design it for LLMs back when models like Qwen 3.6 27B or 3-bit KV cache techniques weren’t the standard. Yet, savvy users are flocking to it because it outperforms the Pro series (4000 through 6000) in real-world efficiency.
Smart users don’t just follow manufacturer guidelines; they adapt to innovation. Expect the demand for the 5090 to skyrocket as the community continues to optimize what is “possible” beyond official branding.
Yes, THAT WAS in the marketing slide. DGX Spark is a inference capacity play, not a inference performance play. Sorry that you missed the memo.
Similarly bragging about RTX 5090 running faster than RTX 6000 or DGX Spark for a single user is clearly missing the point. You can use your tricycle to fetch someone around, doesn’t turn it into a bus. You can race your car on the highway, but that doesn’t turn it into a race car on a professional track. The marketing department doesn’t dictates the use case, they highlight the use case according to the hardware. The product and engineering department dictates the hardware and therefore the use case. If you want to bring your home broadband router into the DC because it’s “faster and cheaper” than the firewall IT bought, that’s your choice, I’m sure you can make an argument about it being a valid use case. My last reply on this topic. Have a good day.
I must say that I’m actually quite happy of my Jetson Thor.
Yeah, this topic is about the DGX Spark, but the Thor is very similar: maybe a little less powerful and with some less features (in some areas, in others actually more).
The main point is that Thor is not as expandable, having the high bandwidth connectivity module at 4x25Gb.
But it’s also a bit cheaper. Also, it also runs much cooler at high loads.
Right now there isn’t much I can’t do that I could on a single DGX Spark.
The low consumption and low noise has made it much more pleasant letting it run 24/7.
This unlocked a whole new world of possibilities!
Your friend has a point. The problem with the second DGX is that once it arrives, you quickly start thinking about buying a switch, more cables, and two more DGXs. It’s a very deep rabbit hole.
So, here’s a pro tip for future SPARK users. Make a plan for the upgrade and time it carefully so your partner or spouse can give you the ConnectX 7 cable as a Christmas gift…
I am quite happy with having two, I feel by making it in cluster mode, we can run Qwen3.5 397B which is a game changer. With single spark, I was in a loop of keeping finding and test models and getting frustrated, the only surprise is 27B, but it is just too slow.
I’m also way happier with two than I was with one, feels very useful now but I was almost ready to return my unit when I only had one. I’ve even ordered two more, and since doing so the price has gone up 35% here so if you haven’t seen price increases yet and you’re on the fence it may be time to pull the trigger. Availability seems poor though, been pushed several times already and not arriving till June now.
After playing and creating with DGX Spark, I deeply understand why big tech companies in the U.S. and China want larger and larger ultra-mega-hyperscale AI data centers.
Even when I have a lot of RAM, I want more…
Even when I have high computing power, I want more…
My mouse cursor is hovering over ORDER button for my 2nd DGX Spark machine, but my wallet and I don’t have much guts yet…