ⓘ Nvidia, edited
Nvidia has expanded DGX Spark's clustering capabilities, allowing users to connect up to eight systems for a combined 1 TB of unified memory. The company has also announced inference optimisations promising up to 40% faster agentic AI workloads on select RTX hardware, alongside improvements for DGX Spark and DGX Station.
Nvidia has announced a series of upgrades for its DGX Spark AI mini PC at GTC Berlin. It comes just days before the previously revealed 64 GB DGX Spark models become available on October 23. According to Nvidia, eight DGX Spark systems with 128 GB of unified memory each can provide 1 TB of memory in total. Similarly, clusters built using the newer 64 GB variants can offer up to 512 GB.
Nvidia says the additional capacity should accommodate larger AI models, longer context windows and multiple AI agents operating simultaneously. The company specifically mentions GLM 5.3 as one of the frontier-class models that can benefit from a 1 TB configuration.
The systems communicate using their built-in ConnectX-7 networking hardware. However, Nvidia says eight-system clusters currently require manual configuration. Its Nvidia Sync utility is scheduled to receive support for eight-system configurations later this month, alongside a new Model Launcher feature. The latter will simplify downloading compatible models, deploying them across connected systems and accessing them through coding tools such as OpenCode.
DGX Sparks also gets up to a 30% speed improvement with Qwen3.8 27B, while DGX Station sees up to 20% higher throughput with GLM 5.3 Flash. Nvidia has also confirmed software-level inference optimisations developed alongside the llama.cpp and vLLM communities. The RTX PRO 6000 Blackwell reportedly completes agentic tasks with Qwen 3.8 Flash Next up to 40% faster, thanks to speculative-decoding improvements.
These figures apply to specific models and workloads, not across-the-board performance increases. Nvidia has not provided corresponding eight-system scaling benchmarks in its announcement, so it remains to be seen how efficiently larger clusters handle demanding workloads.
Lastly, Nvidia is introducing a new collection of memory-optimised AI model checkpoints, including Qwen3.8 Flash Next FP4 and GLM 5.3 Flash FP4 variants. These are designed to reduce memory consumption while leaving more room for KV cache, which is particularly important when running models with longer context windows. The upcoming 64 GB DGX Spark systems, meanwhile, retain the GB10 Grace Blackwell Superchip and start at $4,999. Ironic, considering that's how much the 128 GB variant cost not too long ago.
Anil Ganti - Senior Tech Writer - 3139 articles published on Notebookcheck since 2019
I've been an avid PC gamer since the age of 8. My passion for gaming eventually pushed me towards general tech, and I got my first writing gig at the age of 19. I have a degree in mechanical engineering and have worked in the manufacturing industry and a few other publications like Wccftech before joining Notebookcheck in November 2019. I cover a variety of topics including smartphones, gaming, and computer hardware.