DGX Spark
NVIDIAworkstation| Overview | |
|---|---|
| Architecture | Blackwell (GB10) (2025) |
| Compute capability | 12.1 |
| Memory | |
| VRAM | 128 GB LPDDR5x (unified, shared with the CPU) |
| Memory bandwidth | 273 GB/s |
| Compute — matrix / tensor | |
| FP4 (MXFP4 / NVFP4) | 500 TFLOPS |
| Board & system | |
| TDP | 140 W |
Compatible CUDA Toolkit versions
PyTorch on the DGX Spark
Check against your driver and Python version →
The newest PyTorch, 2.14.0, runs on this card with a plain pip install torch.
| PyTorch | CUDA builds that run on this card | Plain pip install torch |
|---|---|---|
| 2.14.0 | 13.0, 13.2 | works |
| 2.13.0 | 12.9, 13.0, 13.2 | works |
| 2.12.0 | 13.0, 13.2 | works |
| 2.11.0 | 12.8, 12.9, 13.0 | works |
| 2.10.0 | 12.8, 12.9, 13.0 | works |
| 2.9.0 | 12.8, 12.9, 13.0 | works |
| 2.8.0 | 12.8, 12.9 | works |
| 2.7.0 | 12.8 | doesn't (CUDA 12.6 build) |
| 2.6.0 | none | doesn't (CUDA 12.4 build) |
| 2.5.0 | none | doesn't (CUDA 12.4 build) |
| 2.4.0 | none | doesn't (CUDA 12.1 build) |
| 2.3.0 | none | doesn't (CUDA 12.1 build) |
| 2.2.0 | none | doesn't (CUDA 12.1 build) |
| 2.1.0 | none | doesn't (CUDA 12.1 build) |
| 2.0.0 | none | doesn't (CUDA 11.7 build) |
| 1.13.0 | none | doesn't (CUDA 11.7 build) |
| 1.12.0 | none | — |
| 1.11.0 | none | — |
From each build's compiled architecture list (Linux x86_64 wheels). Your driver must also support the CUDA version — see compatible CUDA versions.
Libraries on the DGX Spark
Whether each library's own requirements cover this card. Each row links to the version and source it was checked against on library support.
| Library | Supported | Details |
|---|---|---|
| vLLM 0.30.0 | Yes | |
| FlashAttention-2 2.8.3.post1 | Yes | |
| FlashAttention-3 main (2026-09-27) | No | no build for compute capability 12.1 |
| bitsandbytes 0.50.2 | Yes | |
| TensorRT 11.3.0 | Yes | |
| llama.cpp master (2026-09-27) | Yes |
Native low-precision formats
Whether the card's matrix hardware runs these formats natively. Without native support a format can still work, but through slower emulation or conversion, not at the card's rated speed.
| BF16 | FP8 | FP4 |
|---|---|---|
| Native | Native | Native |
Source: NVIDIA TensorRT 11.3.0 support matrix (hardware precision table).
Added 2026-09-27: a desktop GB10 Grace Blackwell system, popular for running large models locally because of its 128 GB of unified memory. Sources, all NVIDIA's DGX Spark page (nvidia.com/en-us/products/workstations/dgx-spark/): 128 GB LPDDR5x "coherent unified system memory", 256-bit, 273 GB/s; "Up to 1 PFLOP FP4", footnoted as theoretical FP4 with sparsity, so fp4TFLOPS here is the dense half, 500; GB10 TDP 140 W, footnoted as the whole chip (CPU and GPU); 240 W power supply. Compute capability 12.1 from NVIDIA's TensorRT support matrix ("12.1 (DGX Spark)"). vramGB is the full unified memory; the CPU shares it, so less is available to models in practice. The CPU is Arm (20-core, 10 Cortex-X925 + 10 Cortex-A725), so software needs aarch64 (Linux arm64) builds: PyTorch's CUDA 13 aarch64 wheels include Blackwell code, and its RELEASE.md lists 12.0 for aarch64 builds, which covers 12.1.