Edited by 
Short answer
NVIDIA DGX Spark is a palm-sized AI computer built around the GB10 Grace Blackwell Superchip, with 128 GB of unified memory shared by its 20-core Arm CPU and Blackwell GPU. NVIDIA rates it at 1 petaflop of FP4 AI compute (with sparsity) and says it can run models of up to 200 billion parameters, or 405 billion with two units linked. Its weak spot is memory bandwidth, 273 GB/s, so it runs large models that no gaming GPU can hold, but generates tokens more slowly than an RTX 5090.
DGX Spark specifications
| Specification | DGX Spark |
|---|---|
| Chip | NVIDIA GB10 Grace Blackwell Superchip |
| GPU | Blackwell, with 5th-generation Tensor Cores |
| CPU | 20 Arm cores (10 Cortex-X925 + 10 Cortex-A725) |
| AI performance | 1 PFLOP FP4 with sparsity (about 0.5 PFLOP dense) |
| Memory | 128 GB LPDDR5x unified, 256-bit |
| Memory bandwidth | 273 GB/s |
| Storage | 1 TB or 4 TB NVMe |
| Networking | ConnectX-7 SmartNIC, 10 GbE, Wi-Fi 7 |
| Largest model | Up to 200B parameters; 405B with two linked units |
| Size and weight | 150 × 150 × 50.5 mm, 1.2 kg |
| Operating system | NVIDIA DGX OS (Ubuntu Linux) |
Inside the GB10 superchip
GB10 is a scaled-down relative of the GB200 superchip used in data centers. It puts a Grace-style Arm CPU and a Blackwell GPU in one package, joined by NVIDIA's NVLink-C2C chip-to-chip link, so both share the same 128 GB pool of memory. That is the key idea: instead of a GPU with its own small, very fast VRAM, the whole system memory is available to the GPU, like on Apple's Macs. The Blackwell GPU supports the same NVFP4 and FP8 formats as data-center Blackwell chips, and runs the full NVIDIA software stack (CUDA, PyTorch, TensorRT-LLM, vLLM, Ollama).
Which models fit
| Model | Memory needed (4-bit, 8K context) | Fits in 128 GB? |
|---|---|---|
| Llama 3.1 8B, Qwen3 8B | about 6.5 GB | Yes, with room for very long contexts |
| Qwen3 32B | about 24 GB | Yes |
| Llama 3.3 70B | about 50 GB | Yes |
| gpt-oss-120b (MXFP4) | about 69 GB | Yes |
| Qwen3 235B-A22B | about 158 GB | Only with 3-bit quantization, or on two linked units |
| Llama 3.1 405B | about 275 GB | No; two linked units at low precision |
How fast is it for LLMs?
Generating text one token at a time is limited by memory bandwidth, because every token reads every active weight. A useful estimate is about 70% of bandwidth divided by the bytes read per token:
- Llama 3.1 8B in 4-bit (about 4.9 GB): roughly 39 tokens per second.
- Llama 3.3 70B in 4-bit (about 43 GB): roughly 4–5 tokens per second, usable but slow.
- gpt-oss-120b, a mixture-of-experts model with only about 5.1B active parameters: far faster than its size suggests, because each token reads only about 3 GB.
Prompt processing and fine-tuning are compute-bound, and there the Blackwell GPU's Tensor Cores are a real strength compared with other unified-memory machines. In short: DGX Spark is best with mixture-of-experts models and workloads that need lots of memory and CUDA, not for the fastest chat with large dense models.
DGX Spark vs RTX 5090, Mac Studio and Ryzen AI Max
| System | Memory for AI | Bandwidth | Strength |
|---|---|---|---|
| NVIDIA DGX Spark | 128 GB unified | 273 GB/s | Large models plus the full CUDA stack |
| NVIDIA RTX 5090 (desktop GPU) | 32 GB VRAM | 1,792 GB/s | By far the fastest for models that fit |
| Apple Mac with M4 Max | up to 128 GB unified | 546 GB/s | About twice the bandwidth, quiet, macOS |
| Apple Mac Studio M3 Ultra | up to 512 GB unified | 819 GB/s | The largest models on one desktop |
| AMD Ryzen AI Max+ 395 systems | up to 128 GB unified | 256 GB/s | Similar capacity, ROCm/Vulkan software |
Who should buy one
- Good fit: AI developers who want to prototype and fine-tune locally on the same CUDA stack they deploy on in the data center; people who need to load 70B–200B models; teams experimenting with NVFP4.
- Not the best fit: anyone mainly chatting with dense models under 32B (an RTX 4090 or 5090 is much faster), gamers, or Windows users (it runs Linux on Arm).
Frequently asked questions
Is DGX Spark faster than an RTX 5090?
For models that fit in 32 GB, no: the RTX 5090 has about 6.5 times the memory bandwidth, so it generates tokens much faster. DGX Spark wins when a model needs more than 32 GB.
Is DGX Spark the same as Project DIGITS?
Yes. NVIDIA announced the machine as Project DIGITS in January 2025 and renamed it DGX Spark.
Can two DGX Sparks work together?
Yes. Their ConnectX-7 network ports can link two units, which NVIDIA says lets them run models of up to 405 billion parameters.
Check any model on DGX Spark in the LLM VRAM calculator, compare other options in the best GPU for running LLMs locally, or see its data-center sibling in GB200 NVL72 explained.
Leave a Reply