Honestly it’s looking like pre orders are nothing but fluff. Retail stores already have supposed delivery dates up so it looks like inventory is being sent out but what about the preorders?
May → July and now August with no news in site.
10 Likes
I am checking the internet for any news each day, but I am loosing patience.
6 Likes
I am the same as you had to save up and put the money away to get 1 DGX Spark, will soon have enough to get 2 DGX Sparks, probably by the time this actually launches; however, due to Nvidia saying nothing, I am thinking they don’t deserve ny heard earned cash, and maybe I should buy the top of the range fully loaded Mac Studio at £10,000
5 Likes
+1
There is definitely a need for a official update on, what is going on.
Each delayed month without a “real” update on, whats going on is pulling customers in another direction, like a fully loaded Mac Studio.
I am only willing to wait so long, especially, without a update.
Nvidia could do so much better, IMO.
7 Likes
I’ll wait until late August, then I think I’ll opt for the 5090 and a nice PC loaded with RAM. It won’t be integrated but I should be able to achieve the same performance. I’m tired of waiting.
6 Likes
Let’s see whether the T5000 comes out before the Spark. Honestly the segmentation is pretty bad, I guess the T5000 probably has more Tensor Cores and less CUDA cores so it’s more geared towards inference (less FP32 but more TOPS I suppose, not like a 5070 training models is exceptionally good anyway for the Spark).
Might as well get whichever comes out first if either or is not delayed if mostly for inference. T5000 probably is less consumer desktop-friendly but I’ve been waiting ages.
1 Like
Updates from a retailer, take with a grain of salt.
NVIDIA manufacturing delays. Slated for September compared to late August a few weeks ago. No concrete timeline, it’s a fire they’re trying to put out I guess.
1 Like
Thats also what was told, by my supplier
It’s August 21 and still no news about the availability of DGX Spark. No official statement, no estimated date. And the total silence doesn’t sound good at all… If they had at least released a message saying it was postponed until a certain date, that would have been something. But as it stands, everything seems left hanging. I’m starting to believe the project was quietly buried and NVIDIA is shifting its focus elsewhere.
1 Like
I clearly remember that when Jensen introduced the DGX Spark (https://youtu.be/UaGFUtqravQ), he mentioned that it would be “the perfect Christmas present” (at 3:05 of the video above). I think that’s the more realistic expectation :-)
1 Like
Yup. Just had the same email in the UK today.
Seems they have been having issues with the GB10 chip, I think everyone knows they are late to the game (256GB memory is low now). I remember when this first released it was impressive. Interested to see what the actual numbers come out to be upon release. Also interested to see how these perform in a “node” config with the high speed interconnects.
DGX Spark still has a couple of advantages: 1, 256GB total memory with the NVLink set up (of course the price doubles) 2, The DGX OS. Yes, you might be able to get more VRAM and higher bandwidth with a Mac Studio, but it’s probably going to be more expensive and the GPU raw power in terms of TFLOPS (1000 for Spark) isn’t as clear.
1 Like
To be fair, we didn’t place pre-orders since we didn’t pay anything for those “reservations”.
1 Like
Same here, Switzerland
Yep, got the same update this morning.
I was wrong about the TFLOPS: for FP32, Spark only has 80, but 1,750 for FP 16, but a whopping 20 PFLOPS for FP 4 (for quantized models)!:
Apple’s M3 Ultra 80-core GPU delivers approximately 28 TFLOPS of FP32 compute power, based on doubling the FP32 throughput of M3 Max (14 TFLOPS). In contrast, the NVIDIA Blackwell GPU used in DGX Spark achieves 80 TFLOPS of FP32 compute power, making it nearly three times faster in terms of raw FP32 performance.
For lower precision AI workloads, the Blackwell is even further ahead:
-
Blackwell B100 (data center edition) offers ~1,750 TFLOPS of FP16, and up to 20 PFLOPS of FP4 compute with sparsity optimizations.
-
M3 Ultra’s FP16 performance is around 80 TFLOPS, which is only about 4% of Blackwell’s FP16 throughput.
Summary: DGX Spark’s Blackwell GPU strongly outperforms M3 Ultra’s 80-core GPU in raw floating point compute (FP32, FP16, FP4). M3 Ultra’s strengths come from its massive unified memory, energy efficiency, and integration, but in direct TFLOPS comparisons, Blackwell is the clear leader for AI compute tasks.
7 Likes
Super interesting insights ! Thanks for this !
Here in Spain too 😰 No DGX