We’ve just shared an announcement regarding the NVIDIA DGX™ Spark Founders Edition reservations. Details here. https://forums.developer.nvidia.com/t/nvidia-dgx-spark-reservation-update/342874
The email update this morning is a positive sign for communication on product updates, but candidly it also raises more concerns about product quality.
After reserving a DGX Spark in the first minutes of the reservation system going live I’ve been increasingly excited about this new product category. A dedicated companion system for local training, fine tuning, and inference is the perfect compliment to an existing workstation setup.
The challenge comes with expectation setting on the product launch. Since this is a new product offering the delays without an explanation really makes me think there’s an undisclosed product quality issue. Being an early adopter comes with speed bumps along the journey, however at this price point I would hope there will be more transparency on the delays as we get closer to order fulfillment.
Project DIGITS’ announcement in January was very exciting. Pivoting to the DGX Spark product announcement in March was a bit unexpected and generated considerable momentum for interest. The delays on shipping have led to increasing speculation and desires for an alternative product. Candidly I have no interest in Connect-X and would greatly prefer a DGX Spark with 256GB+ and no Connect-X.
I share all of this to vent some frustration but also share support for hopefully increased transparency on the delays as we get closer to launch. In keynotes with Jensen has shared the roadmap for data center product offerings in support of empowering customers to make the right purchasing decisions over time. Hopefully a similar approach can be taken on DGX Spark so that early adopter customers feel informed on their purchasing decisions.
Same email arrived here in Germany!
Same here, today, U.S.
Yep, and then you have of course the CUDA vs Apple Metal debate. Very interesting market out there right now.
Private customers are absolutely sidelined. Vague timelines, no concrete information and always delays… I guess I am cancelling my res
This also applies for reservations of Asus Ascent GX10 via NVIDIA, correct?
I got the message by e-mail, but it only refers to DGX Spark FE.
That’s the decision here isn’t it. Mac Studio with 512GB or 2x Sparks for 256GB. Is Nvidia software worth the premium?
yup, Spark is faster for the models you can fit on it. But that’s the rub isn’t it.
128GB of unified memory at $3k, with the Blackwell GPU seemed incredible when they announced it almost 9 months ago. Now, it seems good but not as amazing as it would have been 9 months ago. The price crept up, and lots of other options for economical local inference have emerged, including the Mac pro with 512GB (LLM tests haven’t been impressive though). Point is, things move fast. Blink and you’ll miss it.
Other then this being obvious AI slop, this is wrong on several levels as its conflating the B100 product with the Spark.
The Spark is not going to outperform a Desktop 5090 while only using LPDDR5x with a narrow 256bit memory bus.
The benchmaxxed FP4 numbers will only be hit when working with data that fits entirely within the on die cache.
While I am disappointed, I am sure they are doing their best. Proud and happy to be part of the AI revolution.
This is already like 2 months overdue with no real reasons nor updated delivery date. NOT PROFESSIONAL! Already thinking that by the time it gets released it will already be outdated.
Just took way to long to only hear a peep!
This sounded good before the wave of AMDs Max+ 395 builds storming the gates. I’ll grab the Corsair Workstation 300 for half the price delivered next month. I have no doubt the Spark will be beautiful but that extra oomfph isn’t worth double for my needs.
I don’t really follow the math of this - but this is not addressing memory bandwidth which is key for inferencing. From what I understand, the Mac a lot faster than the Spark in listed specs.
Is it really the case that for inferencing we should go with high memory bandwidth as a primary performance indicator, and for model generation etc the TFLOPS matters a lot more?
Are you saying the GB10 Superchip with the Blackwell GPU is going to come with 0 build in memory for the GPU? This seems to be what most people might be missing, as NVIDIA has not explicitly announced it. The GB10 is based on the same family architecture as the GB200 or GB300 chips. In the those chips, each Blackwell GPU is designed with on-package HBM3e.
GB200 (2 Blackwell tiles) = 384GB HBM3e → 192GB per tile
GB300 shows similar stats per Blackwell GPU tile.
A GPU running on ONLY LPDDR5x would bottleneck instantly, there is almost no way NVIDIA would build a developer DGX with 0 HBM, it would run slower than a 5090 and worse than an Apple Studio M3 Ultra.
Fact: Every Blackwell GPU SKU has HBM3e; NVIDIA’s own architecture briefs confirm this.
Inference: Spark FE likely carries ~192 GB HBM3e per GPU tile, based on the GB200 split.
Unknown until hardware ships: The exact HBM capacity in Spark FE. Could be 144 GB, 192 GB, maybe 256 GB depending on binning and cost.
I guess we will find out soon enough.
I see this listed in the details for the DGX Station, but not for the DGX Spark.
The DGX Spark datasheet lists ‘Unified Memory’ and a ‘Memory Bandwidth of 273GB/s, with no reference to HBM3e. workstation-datasheet-dgx-spark-gtc25-spring-nvidia-us-3716899-web.pdf
Where was this seen that is applicable to the Spark?
Its not listed for the Spark. The theory lies on the fact that not a single Blackwell GPU has shipped WITHOUT HBM3e memory. Can you prove that wrong? Do you think the Spark is the first Blackwell GPU to not have HBM3E memory?
We can agree to disagree, if that’s your belief. I tend to lean towards them keeping the supply chains clean not not re-designing their whole architecture.
The real question is, why are they leaving out this critical feature in the spec sheet/marketing?