NVIDIA Dual DGX Spark, linked (2× 128 GB, 2× 4 TB)
NVIDIA Dual DGX Spark, linked (2× 128 GB, 2× 4 TB)
Two Sparks for running more agents or models side by side.
- Puppy One fit: Pro (with caveat)
- Memory for models: 2x 128 GB (not one pool for LM Studio)
- Chip / GPU: 2x NVIDIA GB10
- OS: DGX OS (Ubuntu, ARM64)
- Validation: Needs testing: Linux ARM64 (LM Studio officially supports GB10)
- Availability: Available
- Note: LM Studio runs a model on one machine, so the pair does not act as 256 GB for Puppy One.
Two NVIDIA DGX Spark systems linked over their built-in ConnectX-7 SmartNICs at 200 Gb/s, each with 128 GB of unified LPDDR5X memory (2× 128 GB), 40 Arm CPU cores, 8 TB of NVMe storage, and up to 2 PFLOPS of FP4 AI performance across two 150 mm square chassis.
Designed to run and fine-tune models that exceed a single unit's memory, with DGX OS, NVIDIA's AI software stack, CUDA, and NCCL for multi-node inference preinstalled, and the same containers that run on DGX Cloud.
Sits on one desk drawing under 480 W from two standard outlets, with a single 400G QSFP112 direct-attach cable between the units, and runs models up to 405 billion parameters at FP4 across the pair.