even if spark 2 was exactly them same as it is now but with say 800gb/s memory bandwith it would be a huge upgrade
A company with dedicated and specialized GPU/CPU/Memory should be able to get 3X this speed, as they already achieve 2X the 800GB/s with current gen desktop GPUā¦ Rubin platform should be substantially faster, even if the case has to double in size to achieve itā¦
Iād like to see something between the Spark class and the DGX Station class.
Spark is laptop-class power budget and ~$5000, DGX Station is mid-tower class power budget and ~$100kā¦ I want something between those.
Something that would be a clear winner over an M5 or M6 Ultra Mac Studio in terms of power budget, memory, and performance.
So probably:
Vera Rubin
128GB to 192GB GDDR7+ and 256 to 512GB LPDDR6
Power budget around 1000W, same-ish size as Mac Studio
In the $15-20k USD ballpark
To me one of those would be far more useful than a stack of four Spark 2s. Basically cram the Vera Rubin successor to the RTX PRO 6000 Blackwell into a GPU-in-a-box NUC deal.
Iād also like to see a dual-GPU version with double those specs as a single integrated unit.
This would serve SMBs with on-prem requirements well for the class of models emerging as āworkhorsesā, serious enough for real agentic work without going into the long tail of costs for frontier capabilities.
@entrpi something between Spark price and Station price is the NVIDIA GB200 Developer Kit made by Supermicro only. The price for the kit is ~$49K.
See details at Supermicro Introduces Super HPC Station Powered by the NVIDIA GB200 Developer Kit: Data Center Performance in a Deskside Form Factor
Also at NVIDIA GB200 Developer Kit
Thatās still closer to a DGX Station than a Spark. I want something that would be a more practical single-unit choice to someone who would otherwise consider 4 to 8 Sparks or 1 to 2 RTX PRO 6000.
The HBM3e in that GB200 Dev kit is the main differentiator for datacenter-class. GDDR7 or a future GDDR7W would be a good cost-compromise for the missing middle Iām proposing.
If anyone wants an inside-baseball look at what the frontier of local inference will likely look like in 18 months, from someone in the weeds of optimizing the latest and greatest open models for workstation-class CUDA, hereās my research session on that with GPT-5.5 Heavy today:
CORE: H-series level OP CODE support
RAM: At least GDDR-class memory
STORAGE: NVMe 2280 support
OTHERS: Large cooling fans
Now that rtxDgx is announced, whatās your hot take?
Fwiw, Iād like to see the Spark 2 with at least a c2c 2x increase in bandwidth to compete with upcoming Mac m6.
But more realistically, Iād just like to SEE a Spark 2 be announced. Without a second on deck, sparks are going to suffer a tidal wave of grey market collapse.
Looking at the keynote speech today, the DGX Spark replacement will come in 2028 followed by another one in 2030.
I think the next Spark release 2028 will be too late, which means NVIDIA will lose market share due to AMD has already announced the next Gorgon Halo with 192GB memory and Memory Speed of LPDDR5x-8533.
In addition, it seems the AMD Medusa Halo (Strix Halo next) TSMC N2P might have 384-bit LPDDR6 could offer up to 512GB ram (link to AT3 AMD GPU) and likely to launch in 2027.
I hope NVIDIA will have something in place in time.
If you look at the PCB, there isnāt room for much more on the Spark, you will probably see LPDDR6 and not much more besides hardware upgrades. Most of the other things will simply not occur unless thereās some reason for higher density.
Now if you were to use the laptop PCBs, it would be āconceptually easierā to imagine a simple way to double / quadruple memory bandwidth which is simply by 2x/4xāing channels like how the Mac achieves their bandwidth speeds.
GPUDirect support makes no sense. In a sense you already have GPUDirect because GPUDirect is for Storage ā VRAM and since there is no VRAM there is no GPUDirect, the equivalent is DMA (I believe is the umbrella term).
Given the consumer artificial segmentation I donāt think there will ever be a datacenter variant - those will be left to the other product segmentation they have AGX Orion or something.
If itās just going to LPDDR6 memory in the Spark 2, how much of a memory bandwidth gain would that be? Would there be any improvement to the chip itself, say going from GB10 to VR10? Perhaps the memory capacity will be more increased from 128 to 256GB?
They are going to have to do something special, as the newly announced RTX Spark PC line with nearly the exact same Specs as our gen 1 Spark will surely cause the āenthusiast/light enterpriseā market versions to crash and burn.
They even announced a DGX Station with Windows as well. Is this the end of the line for the DGX series, or will it be just relocated to the ālinuxā version :-(
Better customer support. From my experience would avoid a DGX Spark v2 like the plague unless they fix customer support.
They should add screw-terminal connector for external 48V power voltage (and specify exact limits on power inputs), so could be used without usb-c bricks and with batteries.
Where did we get 2028 for the Spark 2? I think it will be October of 2027. It released in October of 2025. 2 years would be 2027. The middle slide says 2027 and 2028. If cadence remains we will see the VR-10 Spark and then 5-8 months later the RTX version. This is cool because it shows the new NIC card the chip and the memory type, LPDDR6. That memory will be faster. It would be nice to get a wider bus still. It doesnt say but moving to SOCAMM would make it expandable. we shall see what happens.
