Spark I lite
Spark v1.5= 50% less ram for same price from 9 months ago.
Dont blame nvidia for ram prices
are you sure?
I don’t get it—5,000 for 64GB times four would be 20,000 more than the M5 Ultra? Or am I missing something? I assume the price of the 128GB version will go up, but how much could it rise compared to the M5 Ultra?
The MLX interfaces on your spark are $1500 on their own
They are integrated into “motherboard”, so they add value to the overall solution but they don’t worth anything on their own, since you can’t use them on their own. You can argue how much value they are adding but they cost nvidia at most $200 in parts, probably much much less.
Oh stop it, do you think it costs nothing because you can’t take it out and use it elsewhere?
Please you are searching extensively for options to satisfy your original hypothesis. When, in acutal fact you probably should have dumped the hypothesis.
Actually, let me think a bit more, … oh I’m sounding like an AI model again.
The $1,500 cost for the MLX in Spark is an obvious forum joke. Does anyone “here” seriously believe that? (I say “here” because this specific forum is where people usually uncover the important nuances of this hardware).
- Nvidia is an absolute pro at selling the exact same thing at different price points (market segmentation for the same chip. An old but classic example: GeForce, Quadro, Tesla).
- The fact that the network card in Spark shares the same “name” as a $1.5k enterprise card doesn’t automatically give it the same value. Why?
Because a chunk of the features found in the $1.5k card is simply missing in Spark:
- Half the speed: 400 vs 200 on a single port, or 200+200 vs 100+100 on dual ports.
- No InfiniBand: Spark only supports RoCE, whereas a standalone ConnectX-7 supports both InfiniBand and Ethernet (this exact issue was already broken down in one of the forum threads).
- No port bifurcation/splitting support: Mellanox supports it natively, but Spark’s implementation does not. This is exactly why a non-ring, switchless architecture for 4-5 Sparks never worked out (also deeply analyzed in other threads on this forum).
There are bound to be other differences too.
It’s strange to ignore all of this and claim, “This is a fully-fledged ConnectX-7 equivalent to a $1.5k card.” No, it’s not. Given the stripped-down functionality, the MLX in Spark is closer to $0.5k than $1.5k. You can argue about the exact numbers, but these are completely different devices in terms of their capabilities.
Well it is probably their retail pricing estimation, not the cost. Cost is probably around 500 or less. But as a point of reference:
https://www.fs.com/eu-en/products/242589.html?attribute=128165&id=4997717
Full fledged 2x200GB/s cx7 card is 2500 eur with vat. Server-grade components ain’t cheap. Priced for enterprise with a massive profit margin.
Hello,
I agree with everyone, doubling the bandwidth would yield the biggest performance gain, making it possible to run 8-bit quantized models. I also feel like 300B to 500B parameters seems to be the sweet spot for self-hosted models. So I would recommend this for the next product if it can achieve both: something like 300 GB of VRAM with double the bandwidth.
That being said, for a different product line, isn’t bandwidth roughly proportional to the memory architecture? When I build a PC, my target memory capacity basically already filters out most other hardware choices. This is even more true with ARM unified memory architectures. Such a system would end up large, more like a medium-sized desktop and might target a completely different market segment. If they want to keep it compact and affordable (< €5k), they might be better off stacking VRAM and improving the LAN network connections instead.
Am I right in assuming that stacking VRAM is easier than increasing bandwidth from a product design standpoint?
I do think the Spark II to competitive, it has to leapfrog and one up its largest competitor, the MAC Studio with 1.2 TB/s. But just getting the Spark II on the same memory BW would at least get it in the same playing field. We know it can be done, as the NVIDIA PCIe product BW speeds have proven they have the ability to have high BW.
Plus, now that the use case has changed from local model training intended use to local inference use, there should albe a priority to fix this massive gap issue taht most on here have requited.
A Vera chipset with 1.7TB/s would be a worthy upgrade!
Training was actually one of my biggest disappointments with the DGX Spark initially, mainly because training throughput is so directly bound by memory bandwidth. Since I don’t train large LLMs but work on embedders and classifiers instead, I just run them on my RTX 5080, it’s honestly at least 8x faster.
I gave the notebooks a shot for LLM RL training and LoRA fine-tuning, but it’s super tricky, and fast iteration times are everything when you’re tweaking setups. Since my workflow relies heavily on rapid iteration and repetition, the DGX Spark just isn’t the right tool for training. In my view, it really should be pitched pure and simple as a dedicated inference unit.
Bottom line: it lacks too much memory bandwidth to serve as a proper MLOps machine, but it excels as an inference engine.
Also, I’ve read comments here and there saying the cooling system is terrible, but my experience has been completely different. I’ve had a Founder’s Edition since November 2025, and I find the thermal design really smart and effective. Overall, the hardware product is solid, 'd argue only the software feels a bit clunky, but thanks to the amazing community, that’s not a dealbreaker for me.
100% agree with you, the community here has pulled a lot a value out of this little box of gold.
There is also a thread with a mention that they are bringing us up to date with current Ubuntu LTS and Cuda (Finally) and hopefully there is deployment mandate to keep us on the bleeding edge of releases!