NVIDIA DGX Spark
CUDA local inference/finetune up to ~200B params. AgentRig Score: price pending
- Chip
- GB10 Grace Blackwell
- CPU cores
- 20
- Memory (GB)
- 128
- Memory type
- LPDDR5x unified
- Max memory (GB)
- 128
- GPU
- Blackwell GPU
- VRAM (GB)
- shared
- Storage
- up to 4TB
- Power
- 240W PSU / 140W TDP
- OS
- DGX OS (Linux)
- Price
- Price pending
- Price config
- DGX Spark 128GB
- Price status
- pending
- Price note
- Specs verified; nvidia.com shows no price
- Llama 3.1 8B Q4 tok/s
- 43.1
- Good for
- CUDA local inference/finetune up to ~200B params
Spec source: https://www.nvidia.com/en-us/products/workstations/dgx-spark/
Price source: https://www.nvidia.com/en-us/products/workstations/dgx-spark/
Benchmark source: https://devquasar.com/uncategorized/gpu-gguf-inference-comparison/