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.
DGX Spark is a compact AI workstation launched by NVIDIA for individual developers and small teams, featuring the Grace Blackwell chip.
The NVIDIA DGX Spark is equipped with the GB10 Grace Blackwell Superchip and 128 GB of unified memory.
Timeline (last 90 days)
This version includes a CI fix: it corrects the outdated NVFP4 decode test selector on the DGX Spark platform (PR #64107).
The NVFP4 decode path on DGX Spark differs from that on data-center GPUs, so it requires separate test configurations and test-case selection logic.
For developers deploying MoE on DGX Spark, it is recommended to introduce optimizations in the order of quantization first, then expert caching and framework-level fusion, to avoid changing multiple variables at once and making it hard to evaluate the gains.
The memory bandwidth of DGX Spark is lower than that of data center GPUs.
The NVIDIA team adopted a batch size of 1 inference setting when evaluating Inclusion AI's Ling-3.0-flash.
The speaker's personal AI assistant 'Magic' runs on a headless DGX Spark, using the Nemotron 3 Super model (120 billion parameters, only 12 billion active).
NVIDIA is building a tiered lineup of personal AI hardware from desktop to workstation: the DGX Spark is close in size to a Mac Mini and suited for individuals; the DGX Station is a game-console-sized workstation that can run trillion-parameter-class models and can be shared by teams.
An 8-node cluster can split a model across different nodes via tensor parallelism or pipeline parallelism, breaking through the memory capacity limit of a single machine.
The 8-node setup suits teams with clear multi-model needs, not all developers.
The NVIDIA developer blog showcased an experiment: using a cluster of 8 DGX Spark nodes to run frontier-level large models locally.
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%VerifiedDGX Spark is a compact AI workstation launched by NVIDIA for individual developers and small teams, featuring the Grace Blackwell chip.
65%VerifiedThe NVIDIA DGX Spark is equipped with the GB10 Grace Blackwell Superchip and 128 GB of unified memory.
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%VerifiedThe combined unified memory of two DGX Sparks totals approximately 256 GB.
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%VerifiedNVIDIA introduced a piece of software called PAIR, designed to aggregate the idle computing power of multiple devices on a home network into a local AI cluster.
65%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%UnverifiedOn a dual AGX Orin system, Qwen 3.8 Flash Next achieves an inference speed of 38 tokens/second.
60%UnverifiedThe interconnect bandwidth between DGX Sparks is expected to be in the range of 100–400 Gbps, far lower than the 900 GB/s bandwidth provided by NVLink within data-center DGX systems.
60%UnverifiedThe NVFP4 decode path on DGX Spark differs from that on data-center GPUs, so it requires separate test configurations and test-case selection logic.
50%UnverifiedFor developers deploying MoE on DGX Spark, it is recommended to introduce optimizations in the order of quantization first, then expert caching and framework-level fusion, to avoid changing multiple variables at once and making it hard to evaluate the gains.
50%UnverifiedThe memory bandwidth of DGX Spark is lower than that of data center GPUs.
50%UnverifiedThe NVIDIA team adopted a batch size of 1 inference setting when evaluating Inclusion AI's Ling-3.0-flash.
50%UnverifiedThe speaker's personal AI assistant 'Magic' runs on a headless DGX Spark, using the Nemotron 3 Super model (120 billion parameters, only 12 billion active).
50%UnverifiedNVIDIA is building a tiered lineup of personal AI hardware from desktop to workstation: the DGX Spark is close in size to a Mac Mini and suited for individuals; the DGX Station is a game-console-sized workstation that can run trillion-parameter-class models and can be shared by teams.
50%UnverifiedAn 8-node cluster can split a model across different nodes via tensor parallelism or pipeline parallelism, breaking through the memory capacity limit of a single machine.
50%UnverifiedThe 8-node setup suits teams with clear multi-model needs, not all developers.
50%UnverifiedDeepInfra's kernels are written for NVIDIA data center Blackwell chips (such as the B200) and cannot be used out of the box on devices using different Blackwell chips, such as the DGX Spark.
50%UnverifiedIn the TP4 configuration on 4 DGX Spark units, measured TTFT is: about 12 seconds for over 60,000 tokens (about 5400 Token/s), about 54 seconds for 260,000 tokens (about 4800 Token/s), and about 4 minutes 54 seconds for 1 million tokens (about 3400 Token/s).
50%