DGX Spark formerly Digits are now available to order at 3,999 for the Nvidia Founders edition and $2999 for the Asus branded Spark named as Ascent GX10. Asus one comes with 1TB whereas Nvidia branded one comes with 4TB NVME space, all models will come with Ubuntu based DGX OS preinstalled. The problem with Spark is that it comes with a meager 273 GB/s memory bandwidth, it is like a 4060ti speed in LLM tasks.
Spark vs AGX Orin:
Spark offers 1.33x greater memory bandwidth (273 GB/s versus 204.8 GB/s), native FP8 support, and double the unified memory.
NVIDIA DGX Spark Digits Official Specifications
| Architecture | NVIDIA Grace Blackwell |
| GPU | Blackwell Architecture |
| CPU | 20 core Arm, 10 Cortex-X925 + 10 Cortex-A725 Arm |
| CUDA Cores | Blackwell Generation |
| Tensor Cores | 5th Generation |
| RT Cores | 4th Generation |
| Tensor Performance1 | 1000 AI TOPS |
| System Memory | 128 GB LPDDR5x, unified system memory |
| Memory Interface | 256-bit |
| Memory Bandwidth | 273 GB/s |
| Storage | 1 or 4 TB NVME.M2 with self-encryption |
| USB | 4x USB 4 TypeC (up to 40Gb/s) |
| Ethernet | 1x RJ-45 connector 10 GbE |
| NIC | ConnectX-7 Smart NIC |
| Wi-Fi | WiFi 7 |
| Bluetooth | BT 5.3 |
| Audio-output | HDMI multichannel audio output |
| Power Consumption | 170W |
| Display Connectors | 1x HDMI 2.1a |
| NVENC | NVDEC | 1x | 1x |
| OS | NVIDIA DGX OS based on Ubuntu |
| System Dimensions | 150 mm L x 150 mm W x 50.5 mm H |
| System Weight | 1.2 kg |
Spark big brother is also launched today, named DGX Station.
Update:
NVidia Spark Inference Speed Token/Seconds
If you wish to use 32GB model q4km like this Deepseek R1 then the file size of that model is 19.9GB, we know the memory bandwidth of Spark is 273 GB/s, so theoretical max tok/s will be around 13.7 tok/second, in real testing this will be a bit lower. So there you have it, the inference speed of Spark (Digits).
Would you buy a $3000 product to get ~13 tok/s?
The higher models like the llama 70B q4km is ~42GB and max inference speed on Spark will be ~6.5 tok/s. When you add some context, it will work at crawling speed 4-5 tok/s if you are lucky.
