Eight DGX Spark nodes teamed with ConnectX‑7 RoCE are now powering frontier‑class AI, unified memory & multi‑agent workloads. See how NVIDIA Sync, inference and clustering boost performance. #DGXSpark #ConnectX7 #FrontierAI
🔗 aidailypost.com/news/8-nvidi...
🔗 aidailypost.com/news/8-nvidi...
October 22, 2026 at 4:41 PM
Eight DGX Spark nodes teamed with ConnectX‑7 RoCE are now powering frontier‑class AI, unified memory & multi‑agent workloads. See how NVIDIA Sync, inference and clustering boost performance. #DGXSpark #ConnectX7 #FrontierAI
🔗 aidailypost.com/news/8-nvidi...
🔗 aidailypost.com/news/8-nvidi...
Getting 4 DGX Sparks talking to each other is going to cost more than this beast because they've only got the one connectx7 port on the back. So like as much as I'm clutching my pearls at the price, there's a niche
August 25, 2026 at 5:08 PM
Getting 4 DGX Sparks talking to each other is going to cost more than this beast because they've only got the one connectx7 port on the back. So like as much as I'm clutching my pearls at the price, there's a niche
📰 ASUS Umumkan Ketersediaan AI Superkomputer Personal Ascent GX10
👉 Baca artikel lengkap di sini: https://ahmandonk.com/2025/10/14/asus-ascent-gx10-ai-supercomputer/
#aiwor#aiworkstationn#ascentgx10 #asuse#connectx7s#datascience #gb10e#graceblackwelll#localaii#nvidiao#personalaisupercomputer
👉 Baca artikel lengkap di sini: https://ahmandonk.com/2025/10/14/asus-ascent-gx10-ai-supercomputer/
#aiwor#aiworkstationn#ascentgx10 #asuse#connectx7s#datascience #gb10e#graceblackwelll#localaii#nvidiao#personalaisupercomputer
October 14, 2025 at 10:02 AM
📰 ASUS Umumkan Ketersediaan AI Superkomputer Personal Ascent GX10
👉 Baca artikel lengkap di sini: https://ahmandonk.com/2025/10/14/asus-ascent-gx10-ai-supercomputer/
#aiwor#aiworkstationn#ascentgx10 #asuse#connectx7s#datascience #gb10e#graceblackwelll#localaii#nvidiao#personalaisupercomputer
👉 Baca artikel lengkap di sini: https://ahmandonk.com/2025/10/14/asus-ascent-gx10-ai-supercomputer/
#aiwor#aiworkstationn#ascentgx10 #asuse#connectx7s#datascience #gb10e#graceblackwelll#localaii#nvidiao#personalaisupercomputer
I only have 1 but I think you can only connect 2 at a time. It has a Connectx7 2 port connection
I haven’t tried any image/video rendering. GMOME was crashing when using ollama and AI studio doesn’t have an arm Linux build
I haven’t tried any image/video rendering. GMOME was crashing when using ollama and AI studio doesn’t have an arm Linux build
October 14, 2025 at 3:07 AM
I only have 1 but I think you can only connect 2 at a time. It has a Connectx7 2 port connection
I haven’t tried any image/video rendering. GMOME was crashing when using ollama and AI studio doesn’t have an arm Linux build
I haven’t tried any image/video rendering. GMOME was crashing when using ollama and AI studio doesn’t have an arm Linux build
Price of the DGX Spark isn't that bad compared to AI Max+ 395 these days. Problem is the router for ConnectX7 is going to be more than 4 Asus GX10 boxes, but you can find a 100 GbE fiberoptic router that won't break the bank. All for around the price of a used Honda Civic.
You can now run Kimi K2.7 Code locally! 🌘
We shrank the 1T model to 325GB (-48%) via Dynamic 2-bit where important layers are upcasted.
Run at >40 tok/s on 330GB RAM/VRAM setups.
Run full precision on 610 GB.
Guide: unsloth.ai/docs/models/...
GGUF: huggingface.co/unsloth/Kimi...
We shrank the 1T model to 325GB (-48%) via Dynamic 2-bit where important layers are upcasted.
Run at >40 tok/s on 330GB RAM/VRAM setups.
Run full precision on 610 GB.
Guide: unsloth.ai/docs/models/...
GGUF: huggingface.co/unsloth/Kimi...
June 16, 2026 at 1:50 AM
Price of the DGX Spark isn't that bad compared to AI Max+ 395 these days. Problem is the router for ConnectX7 is going to be more than 4 Asus GX10 boxes, but you can find a 100 GbE fiberoptic router that won't break the bank. All for around the price of a used Honda Civic.
AFAIK very similar in performance but big deal [I think? don't have direct experience so caveat lector] is nvidia and cuda Just Work out of box w/system memory and I find myself having to fall back to vulkan instead of rocm. That + connectx7 make people willing to pay $1k more? not me!
December 7, 2025 at 8:07 AM
AFAIK very similar in performance but big deal [I think? don't have direct experience so caveat lector] is nvidia and cuda Just Work out of box w/system memory and I find myself having to fall back to vulkan instead of rocm. That + connectx7 make people willing to pay $1k more? not me!
I wish they had the AMD equivalent of a ConnectX7 port on the back of them so I could distribute a model across multiple nodes. I'd pay more.
So @frame.work has the AMD Ryzen AI Max+ 395 and I'm tempted.
August 31, 2025 at 2:39 AM
I wish they had the AMD equivalent of a ConnectX7 port on the back of them so I could distribute a model across multiple nodes. I'd pay more.
Two port ConnectX7 NIC is well north of $1k by itself. That's what makes me still interested vs AMD Ryzen AI Max+ 395. If I only wanted one, I'd get the AMD. But it doesn't have a fast connect like the DGX
August 29, 2025 at 10:01 AM
Two port ConnectX7 NIC is well north of $1k by itself. That's what makes me still interested vs AMD Ryzen AI Max+ 395. If I only wanted one, I'd get the AMD. But it doesn't have a fast connect like the DGX
Especially as you stack them. You can run GLM-4.5 on two DGXs. They have a ConnectX7 port on the back. Equivalent HW setup for 5090s is like $25k, minimum. Can the $25k process more tokens per $? Probably! I'm sure they're much faster, but I'm WAY more likely to buy two DGX clones than a $25k rig.
August 29, 2025 at 9:59 AM
Especially as you stack them. You can run GLM-4.5 on two DGXs. They have a ConnectX7 port on the back. Equivalent HW setup for 5090s is like $25k, minimum. Can the $25k process more tokens per $? Probably! I'm sure they're much faster, but I'm WAY more likely to buy two DGX clones than a $25k rig.