The DGX Spark runs any open model whose weights and context cache fit in its 128 GB of unified memory, so it can hold models that a 24 GB or 32 GB graphics card cannot. It is slower per token than a high-end GPU, because its memory bandwidth is much lower. The fit list on this page shows which models in our data fit.
What the DGX Spark is
NVIDIA builds the DGX Spark, a desktop system, around the GB10 Grace Blackwell Superchip. The specification table lists a 20-core Arm CPU, 128 GB of LPDDR5x coherent unified system memory on a 256-bit interface, 273 GB/s of memory bandwidth, and up to 1 petaFLOP at FP4 precision with sparsity. The power supply is rated at 240 W and the GB10 chip has a 140 W TDP. The box measures 150 mm by 150 mm by 50.5 mm and weighs 1.2 kg. A 64 GB option is listed as coming soon, through OEM partners only.
How unified memory changes what fits
A normal graphics card has its own video memory, and the model must fit in it. On the DGX Spark the GPU shares system DRAM with the CPU. NVIDIA’s user guide says this lets larger amounts of memory be used for GPU workloads, and warns about the side effects. nvidia-smi reports “Memory-Usage: Not Supported” because the GPU has no dedicated framebuffer, and cudaMemGetInfo can report less than the memory that is really allocatable, because the CPU can free pages by swapping.
Plan for three things:
- The operating system and your other programs draw on the same pool as the model. The usable amount is below the headline figure.
- Weights are only part of the bill. The key-value cache grows with context length, as NVIDIA’s inference article explains, so long contexts take memory that short chats do not.
- Quantised weights matter. Fewer bits per weight shrinks the files and leaves room for context.
Mixture-of-experts models suit this machine. The gpt-oss-120b model card lists 117B parameters with 5.1B active, post-trained with MXFP4 quantisation of its mixture-of-experts weights so that it runs on a single 80 GB GPU. That is below the Spark’s 128 GB.
Speed: capacity is not bandwidth
NVIDIA’s inference article says model execution is frequently memory-bandwidth bound, especially in the weights. The DGX Spark reads memory at 273 GB/s. NVIDIA’s compare page lists the GeForce RTX 5090 at 1792 GB/sec, with 32 GB. The Spark trades speed for capacity, and models that read fewer active parameters per token, such as mixture-of-experts designs, are the natural match. A large dense model that fits will run, but expect slower generation than on a card with several times the bandwidth.
Software support
Ollama’s hardware table lists the GB10 in the DGX Spark. NVIDIA’s user guide points to NGC containers for PyTorch, vLLM and TensorRT-LLM. llama.cpp has CUDA support, so GGUF files work too.
Clustering
NVIDIA says ConnectX networking lets you connect up to four DGX Spark systems to work with larger models, faster inference and multi-agent workloads. The specification lists a ConnectX-7 NIC at 200 Gbps, and NVIDIA’s cluster table shows 256 GB for two systems and 512 GB for four. Systems linked this way are separate machines on a network, so expect to set up cluster software.
When renting beats buying
Rent a cloud GPU when:
- you want to test a model before committing to hardware;
- your workload is occasional, so a machine would sit idle;
- you need more memory or more bandwidth than the Spark offers;
- you need a short burst of many GPUs, for example for training.
Buying fits steady, always-on use, such as local agents, or work that must stay on your own machine. NVIDIA’s product page pitches the Spark for always-on agent workloads. We do not quote rental rates, so check current ones with providers.
FAQ
How much memory does the DGX Spark have?
The standard system has 128 GB of LPDDR5x unified memory shared by the CPU and GPU. NVIDIA lists a 64 GB option through OEM partners only.
Can I connect two DGX Sparks?
Yes. NVIDIA documents connecting up to four systems over ConnectX networking, and its table lists 256 GB for two.
Is the DGX Spark faster than an RTX 5090?
Not for token generation on a model that fits both. NVIDIA lists 273 GB/s for the Spark and 1792 GB/sec for the 5090, and generation is often bandwidth bound.
Does Ollama work on the DGX Spark?
Yes. Ollama’s hardware docs list the GB10 in the DGX Spark among supported NVIDIA GPUs.