Personally, I would love to see in v2:
* Rubin platform.
* Multiple options of up to 1TB of RAM, allowing plenty space to run the largest of models locally.
* Better heat management.
* Full memory bandwidth, maybe 25% of Rubin platform at least? 273 GB/s is just painful.
* NemoClaw baked in.
Memory bandwidth is much important than capacity for me now.
1TB is impossible
Painfully slow! My PCIe 4.0 RTX A5000 card from 5 years ago is much more BW.
Added to the wish listā¦.
I did see one person in Youtube link 8 together to get 1TB.
Your wishlist for a Spark v2 is basically the DGX Station: NVIDIA DGX Station
- At least twice memory bandwidth.
- Proper ConnectX 7 interface (not split into two subinterfaces).
- GPU Direct support.
- Working MMAP (if they solve GPU Direct, it will likely be solved too).
- No weird PD/OOM issues.
- Would be nice to have a stripped down datacenter chip vs. RTX variant.
- Power indicator light.
- BMC controller would be nice.
I think people are missing the point of the GB10. If you want more power there are plenty of solutions including the DGX Station that offer way more power. I think the power of the GB10 is the higher RAM and reasonable compute in a friendly package and power envelope that you can run at home. Nvidia sells all kinds of professional solutions I couldnāt dream of running at home.
My wishlist
- Some more memory bandwidth, DDR6
- ConnectX configuration where I can get to 4 nodes without an expensive switch
- Internal power supply with AC input (itās fine if itās physically bigger)
- Power LED, and front power button
- At least USB4
Realistic take:
- Generational bump in architecture
- Architecture parity with server chips
- Memory bandwidth roughly doubled is essential (or high risk against Apple et al) but will not be HBM
- Option for 256GB would be ideal, but if the DGX Station lands around there then 192GB might be the chosen compromise - with high risk against Appleās larger pools
- Ability to swap in multiple sizes of NVME drive, up to 2280
- Thunderbolt port compatibility with Nvidia eGPUs to allow an external but directly attached 3090/4090 to draft or host a smaller model for optimal speed or fancier agentic work.
All that said, I hope this platform is fully polished and fully leveraged with NVFP4 etc. before effort is spent toward gen2.
People fussed when the price was increased from $3,999 to $4,699. All the above wishes will definitely rise the price a lot more.
- NVFP4 fully implemented and performance as promised as a baseline.
- A DGX Spark without ConnextX-7, with the same current memory bandwidth, ram GB and and lower entry price to make AI research more accessible.
- One DGX Sparks without ConnectX-7 and instead with the double memory bandwidth as currently GB/s and double of the memory like 256 GB ram for the same price as the current DGX Spark.
- 2280 NVME for better thermal management.
- And of course the current DGX Spark clusters with ConnectX-7, etc.
Spark ā Ember ā Flame ā Fire ā Blaze ā Inferno
So, please, no Spark II ā
Iām building the Ember, and Iām bringing:
- 3x memory bandwidth (solves a lot: more speed, more time for CPU and RoCE, which also block unified memory access)
- 2x compute (reduces memory-blocking time ā and more is needed because hardware features were removed)
- Real RDMA for RoCE (only this gives significant speedup compared to the RoCE-CPU-path of the Spark)
and Iām leaving at home:
- NVFP4 we donāt need (aside from the new Nemotron natively 4-bit-trained model?), due to reduced accuracy and bad support. we are happy with int4 autoround
And if memory were twice that of the DGX Spark, fewer people would care about a second unit ā although I think that price point would put it out of reach for most individuals entirely.
UPDATE:
- same pricetag or less ;)
Donāt come back here if you bought one and then learned itās āa bit differentā, āit dependsā, or āafter all it doesnāt do what a datacenter GPU does.ā Maybe someone will open a thread asking for a DGX Station Two ā and gets the wishes you had in mind when you bought the first one.
@flash3 the DGX Station forum is up already! Since there are no DGX Stations available yet thereās no traffic on that forum. The new system is expected to be unveiled next week during the Nvidia GTC event.
Based on prices from CDW and MSI the new DGX Station will cost as much as a dozen leather jackets the Nvidia CEO likes to wear.
so I can buy one and pay in leather instead?
Rubin Architecture.
Memory moved to Nvidia standard SOCAMM2, possibly moved to LPDDR6.
The bus likely gets bumped to 384 or 512.
128 gb. Maybe a way to get it to 256 or 512.
Will use N2 chip.
Socamm2and larger bus make it all worthwhile, IMO.
What it boils down to is if the chip N2 can utilize above 128.
I use the DGX Spark primarily for development rather than for service deployment, so compatibility and stability are far more important to me than raw speed. Because of this, I believe Spark 2 should include the following improvements:
- A selectable memory capacity of at least 64 GB up to 1 TB.
- An architecture that is fully compatible with workstation and consumer GPUs, with the same level of support. (Currently, I believe SM121 has limited compatibility and support.)
- High durability, strong stability, and fast support when issues occur, especially for the shutdown problems that have been reported frequently recently.
- I understand that an ARM-based CPU may have been chosen for power efficiency, but even if it consumes more power, I would strongly prefer an x86 CPU for better compatibility.
keep in mind ā¦ GPU, board and memory will be fused, no more lego style.
No one will build different memory setups for this niche.
Fisherās wife was sitting in her hovel again in the end, if you know the story of the Brothers Grimm.
havenāt you heard about the hot hack yet?
In the BIOS, press Ctrl + Super + Alt + N ā a popup appears where you can change the āJensen factorā. The manufacturer default is set to 1x, just change it to 2x or 4x.
The highest setting is very critical for the board ā high memory frequency, overclocked CPU, enabled hidden CUDA cores draw more power. Factor 3 was not included for marketing reasons ā it sounds too much like a number from a fairy tale. So the recommendation is x2.
Itās then so fast that all datacenter planners would immediately cancel their H200 orders and the DGX Spark price would explode. Therefore a password was built in to activate the new factor. You can find it on April 1st in the forum under the āwhy this is not possibleā thread.
Not really, I do expect V2 to be in the same form factor :)
I did that with to a Turbo in a car once, was Epic for 38 seconds!