DGX Spark
A desktop AI all-in-one machine released by NVIDIA, pre-installed with the software stack required for AI inference. It is designed for out-of-the-box use and targets enterprise scenarios requiring rapid AI application deployment.
Core Facts
The CPU and GPU of DGX Spark share up to 128 GB of LPDDR5X memory, achieving memory-coherent access via NVLink-C2C high-speed interconnect.
The NVIDIA DGX Spark is equipped with the GB10 Grace Blackwell Superchip and 128 GB of unified memory.
The combined unified memory of two DGX Sparks totals approximately 256 GB.
Timeline (last 90 days)
In a harness with terminal access to a local DGX Spark GPU node, Argon first reproduced a research paper on growing neural cellular automata, implemented an exact 8,000-parameter model in PyTorch, trained it on the GPU, cut off half of the organism's body to verify self-healing, and then ported the trained weights to an interactive WebGL shader, taking 23 minutes total from paper to a usable browser demo, with the browser shader and the PyTorch model matching to five decimal places.
In the demo, TabFM runs locally on a DGX Spark, with data stored in a Neon-based Serverless Postgres.
In image classification tasks, OpenAI completes in approximately 3.4 seconds, while the locally running Ollama solution with Qwen flash takes approximately 9.5 seconds (running on a Dell Pro Max/GB10 DGX Spark device).
EvoX Agent can connect to LM Studio, Ollama, and standalone devices such as DGX Spark via the AI services page to run local models such as Nemotron and Qwen.
NVIDIA's officially supported maximum interconnect configuration for DGX Spark at the time was 3 units.
A user runs the 397B model on 4× ASUS GB10 units achieving 30 tokens/s at only 400W of power consumption, compared to 50–60 tokens/s at 6kW previously.
The author ran the demo on a DGX Spark; the CLM server itself is approximately 75 MB, the overall footprint is approximately 25 GB of VRAM, and after caching, the same ticket can be queried again with a response within 2 milliseconds.
NVIDIA announced a new 64 GB unified memory version of DGX Spark, available from October 23 through suppliers including Redflag, ASUS, Dell, Gigabyte, HP, and MSI, starting at $4,999, retaining the GB10 Grace Blackwell chip.
Unoptimized baseline token decoding speeds for running Qwen 3 (27B): RTX 5090 ~75, RTX 4090 ~46, RTX 3090 ~40, M5 Max 31.4, DGX Spark 12–17, Ryzen AI Max 11.5 tokens/s.
Cosmos 3 is available in three sizes—Super, Nano, and Edge—where Super pursues the highest fidelity and Edge is optimized for edge devices such as Jetson Thor, Orin, and DGX Spark.
40 more timeline events
All Facts (20)
The CPU and GPU of DGX Spark share up to 128 GB of LPDDR5X memory, achieving memory-coherent access via NVLink-C2C high-speed interconnect.
80%VerifiedThe NVIDIA DGX Spark is equipped with the GB10 Grace Blackwell Superchip and 128 GB of unified memory.
65%VerifiedThe combined unified memory of two DGX Sparks totals approximately 256 GB.
65%VerifiedThe DGX Spark is equipped with the GB10 Grace Blackwell Superchip, 128 GB of unified memory (LPDDR5x), and officially delivers up to 1 PetaFLOP (FP4 precision) of AI compute.
65%VerifiedDGX Spark is a desktop-level AI supercomputer released by NVIDIA at the 2025 GTC conference, featuring the GB10 superchip based on the Grace Blackwell architecture.
80%UnverifiedDGX Spark基于NVIDIA的Grace Blackwell架构，采用Grace CPU与Blackwell GPU的紧密耦合设计
90%UnverifiedDGX Spark is a compact AI workstation launched by NVIDIA for individual developers and small teams, featuring the Grace Blackwell chip.
60%UnverifiedOn a dual AGX Orin system, Qwen 3.8 Flash Next achieves an inference speed of 38 tokens/second.
60%UnverifiedIn a harness with terminal access to a local DGX Spark GPU node, Argon first reproduced a research paper on growing neural cellular automata, implemented an exact 8,000-parameter model in PyTorch, trained it on the GPU, cut off half of the organism's body to verify self-healing, and then ported the trained weights to an interactive WebGL shader, taking 23 minutes total from paper to a usable browser demo, with the browser shader and the PyTorch model matching to five decimal places.
50%UnverifiedIn the demo, TabFM runs locally on a DGX Spark, with data stored in a Neon-based Serverless Postgres.
50%UnverifiedIn image classification tasks, OpenAI completes in approximately 3.4 seconds, while the locally running Ollama solution with Qwen flash takes approximately 9.5 seconds (running on a Dell Pro Max/GB10 DGX Spark device).
50%UnverifiedEvoX Agent can connect to LM Studio, Ollama, and standalone devices such as DGX Spark via the AI services page to run local models such as Nemotron and Qwen.
50%UnverifiedNVIDIA's officially supported maximum interconnect configuration for DGX Spark at the time was 3 units.
50%UnverifiedA user runs the 397B model on 4× ASUS GB10 units achieving 30 tokens/s at only 400W of power consumption, compared to 50–60 tokens/s at 6kW previously.
50%UnverifiedThe author ran the demo on a DGX Spark; the CLM server itself is approximately 75 MB, the overall footprint is approximately 25 GB of VRAM, and after caching, the same ticket can be queried again with a response within 2 milliseconds.
50%UnverifiedUnoptimized baseline token decoding speeds for running Qwen 3 (27B): RTX 5090 ~75, RTX 4090 ~46, RTX 3090 ~40, M5 Max 31.4, DGX Spark 12–17, Ryzen AI Max 11.5 tokens/s.
50%UnverifiedIn Path C, an RTX 3080 paired with a DGX Spark improves Prefill speed from 1111.3 to 1600 (approximately 44%), but Decode speed drops from 33.36 token/s to 17.62 token/s.
50%UnverifiedA six-device configuration of four DGX Sparks and two RTX 6000Ds can locally deploy DeepSeek V4.1 Flash, with a total cost of approximately $38,000 (around 270,000 RMB).
50%UnverifiedUP host 小蛋糕's final decision is not to buy the M5 Ultra and to keep the Spark, primarily because of the need to run AI experiments and because the CUDA platform libraries are ready to use out of the box.
50%UnverifiedThe M5 Ultra's BF16 compute performance is equivalent to only half of a DGX Spark.
50%