RTX Spark vs DGX Spark: specs, VRAM and what differs between NVIDIA’s two Spark computers

RTX Spark vs DGX Spark: the short answer

RTX Spark is NVIDIA’s Arm-based platform for laptops and compact desktops made by several PC manufacturers, with up to 128 GB of unified memory and Windows 11 as the operating system NVIDIA lists, for the person at the desk. The DGX Spark shares the name and much of the published silicon, but NVIDIA builds it for a different job. It is a Linux AI development system with 128 GB of unified memory at 273 GB/s, NVIDIA DGX OS, NVIDIA’s AI software stack and ConnectX-7 networking at 200 Gbps for linking several systems.

NVIDIA lists both chips at CUDA compute capability 12.1. In their largest versions both have 6,144 CUDA cores, a 20-core Arm CPU and a headline figure of 1 petaflop of AI compute. They differ in the operating system, the memory bandwidth, the networking and the way they are sold. This comparison rests on NVIDIA’s pages as read on 16 October 2026. Our DGX Spark specs article has the full DGX Spark data sheet.

For a company team the DGX Spark is the better choice: it runs the same Linux and CUDA stack as the GPU servers the team’s models will move to, and two or four systems link into 256 or 512 GB. An RTX Spark PC suits one person who wants local AI on a portable or desktop computer.

What NVIDIA announced for RTX Spark, and when

NVIDIA announced RTX Spark in a press release on 31 May 2026: a superchip with a Blackwell RTX GPU, connected over NVLink-C2C to a 20-core Grace CPU, for laptops and compact desktops. On 16 October 2026 NVIDIA announced that “RTX Spark laptops are available globally today” from six PC manufacturers. NVIDIA’s developer blog of the same day says the small form factor desktops are “coming in November”.

NVIDIA’s specification page names the chip RTX Spark N1X and lists three versions, two for laptops and one for desktops. NVIDIA’s porting guide for RTX Spark (version 0.1.0, last updated 10 September 2026) describes it as an Arm-based system-on-chip for laptops. In its largest version the CPU has 20 Arm v9.2 cores in two clusters of 10, each with five Cortex-X925 and five Cortex-A725 cores. NVIDIA gives the DGX Spark the same mix of ten Cortex-X925 and ten Cortex-A725 cores.

We found no NVIDIA page that calls the N1X a GB10. NVIDIA’s CUDA compute capability list names them as two entries, “NVIDIA GB10 (DGX Spark)” and “NVIDIA RTX Spark N1X”, both at 12.1.

RTX Spark specs next to DGX Spark

NVIDIA DGX Spark product page (updated 7 October 2026) and hardware overview (10 September 2026); NVIDIA RTX Spark product and specification pages, read on 16 October 2026; RTX Spark porting guide 0.1.0 (10 September 2026); NVIDIA CUDA GPU compute capability list.

The GPU, CPU and compute figures of the largest RTX Spark are the same as those of the DGX Spark. NVIDIA gives the laptop chip a TDP of 45 to 80 W and the desktop chip 140 W, the figure it gives the GB10 in the DGX Spark. For the DGX Spark’s 1 PFLOP, NVIDIA’s footnote reads “Theoretical FP4 TOPS using the sparsity feature.” NVIDIA’s RTX Spark pages carry no such footnote, and StorageReview reported on 7 October 2026 that one PC maker’s footnote describes it as theoretical FP4 performance using sparsity.

For RTX Spark PCs, NVIDIA has not published storage, network ports, size, weight or power supply rating; these are set by each PC maker. The DGX Spark Founders Edition has one fixed configuration with 128 GB and a 4 TB drive.

RTX Spark VRAM and memory speed

Neither system has separate VRAM. On both, the CPU and GPU share one pool of LPDDR5X, and the figure to compare with a graphics card is the unified memory. NVIDIA’s RTX Spark porting guide describes “Up to 128 GB of LPDDR5, shared by the CPU and iGPU” and states that “The memory is fully coherent.” On the DGX Spark, NVIDIA’s playbook settings leave roughly 90 to 115 GB for weights and KV cache, as our specs article works out.

On RTX Spark PCs, the GPU cannot use all of the memory. The porting guide’s chapter on unified memory reserves a carveout for the GPU, which the operating system reports as dedicated GPU memory, although it sits in the same DRAM. NVIDIA does not state the size of the carveout. The memory left after it is split into shared memory, which the CPU and the GPU both use, and CPU-only memory, which “cannot be used for GPU allocations.” The shared part is that remainder minus 16 GB, held between 50 and 80 per cent of it, so the GPU can use the carveout plus at most 80 per cent of the rest.

The memory configuration of an RTX Spark PC is up to its maker. NVIDIA lists up to 128 GB for the 6,144-core versions. For the smaller laptop version with 5,120 cores, its two pages disagree: the product page (updated 16 October 2026) gives “Up to 64 GB LPDDR5X”, the specification page (updated 14 October 2026) “Up to 32 GB LPDDR5X”. Check the memory of the specific PC before planning a model around it.

For memory speed, NVIDIA’s RTX Spark product and specification pages list a 256-bit interface and no bandwidth. The porting guide states: “The interface is 256 bits wide, with a bandwidth of 300 GB/s.” StorageReview calculated 300.8 GB/s from one maker’s listing of LPDDR5X at 9,400 MT/s.

Token generation for one user follows memory bandwidth, because each new token of a dense model reads every weight once. For a dense 70B model with 8-bit weights, about 70.6 GB, the ceiling by our arithmetic is about 3.9 tokens per second at 273 GB/s and about 4.2 at 300 GB/s. LMSYS measured 2.7 tokens per second for Llama 3.1 70B in FP8 on the DGX Spark in October 2025. As of 16 October 2026, NVIDIA has published no tokens per second figure for RTX Spark in the pages we read, and we found no third-party run that states model, engine and date. Our DGX Spark benchmark article collects the measured DGX Spark figures.

Prompt processing depends on compute and, in a laptop, on the power the maker allows within 45 to 80 W. NVIDIA’s RTX Spark product page says it can “Run models up to 120B parameters with 1M context”, while the DGX Spark page gives inference on models up to 200 billion parameters. For more speed per user than either Spark, a workstation card is the next step, as our RTX PRO 5000 72 GB vs DGX Spark comparison shows.

We supply the DGX Spark and workstations built to order with RTX PRO 5000 or RTX PRO 6000 cards. Send us the model and how many people will use it, and we reply within one business day with a configuration and a quote.

Operating system and AI software on each Spark

DGX OS is NVIDIA’s Ubuntu-based Linux for Arm and carries the CUDA stack of NVIDIA’s data-centre GPUs, and the vLLM, TensorRT-LLM and NIM playbooks NVIDIA publishes for the Spark use engines that also run on RTX PRO 6000 and H200 NVL servers. The Founders Edition gets its updates and recovery image from NVIDIA, as our article on buying a DGX Spark in the EU explains.

For RTX Spark PCs, NVIDIA lists Windows 11 as the operating system of all three versions. NVIDIA’s product page says CUDA “runs natively on RTX Spark” and names CUDA, TensorRT and NVFP4 as the stack, and the specification page marks “NVIDIA AI Stack” as supported without listing its parts. NVIDIA’s press release of 16 October 2026 names the first PyTorch-CUDA build for the Arm version of that operating system among the developer tools. Its developer blog points to the Linux subsystem built into the operating system, which gives developers “the same environment and dependencies they will use when deploying to the cloud.” The same blog says Arm builds exist for more than 97 per cent of the top 15,000 Python packages. NVIDIA names llama.cpp, vLLM, Ollama, Unsloth and LM Studio as runtimes for its Nemotron 3.5 Bolt model at launch. In the pages we read, NVIDIA does not name NIM or TensorRT-LLM for RTX Spark.

Both systems have Arm CPUs. On the DGX Spark, an x86 binary-only dependency needs an Arm Linux build. On RTX Spark, NVIDIA’s porting guide says “Many existing x86 and x86_64 applications can run” under the operating system’s emulator, with compatibility depending on the application and its dependencies. The guide adds that emulated code “incurs emulation overhead and performance penalties.” GPU kernels built for compute capability 12.1 target both chips, and the programs around them are built separately for each operating system.

Linking systems and sharing a model with a team

The DGX Spark has two QSFP ports on its ConnectX-7 adapter. NVIDIA’s product page gives two linked systems 256 GB for models of up to 400 billion parameters, and four systems 512 GB for up to 700 billion. Three systems can be cabled directly and four need a switch, as our two-node DGX Spark article describes with the cabling.

NVIDIA’s RTX Spark specification page lists no ConnectX adapter or network interface. NVIDIA PAIR, an open-source tool that NVIDIA moved out of beta on 13 October 2026, lets an RTX Spark PC serve AI inference to other compatible computers on a home network, according to NVIDIA’s press release. NVIDIA’s blog says PAIR “distributes AI workloads across available systems” and does not say that one model can be split across several PCs.

DGX Spark or RTX Spark for a development team

The choice follows from where the team’s code will run in production.

Our reading of the NVIDIA DGX Spark and RTX Spark pages cited above, as of 16 October 2026; RTX PRO 6000 Server Edition from NVIDIA’s product page.

Take a team of 12 developers building an internal assistant that will run on RTX PRO 6000 Server Edition or H200 NVL servers. Two DGX Spark systems give the team a shared test environment on the same Linux and CUDA stack as those servers, each holding a 4-bit model of up to about 200 billion parameters, or one of up to 400 billion across both. Developers with RTX Spark laptops can run smaller models locally, and the Spark systems stay the reference for the server stack. Systems with other unified-memory chips are covered in our DGX Spark vs Mac Studio vs Ryzen AI Max comparison.

We supply the DGX Spark Founders Edition on one EU contract and invoice, with manufacturer warranty. Tell us what your developers build and where it will run in production, and we reply within one business day with a configuration and a quote.

What we supply

We supply the NVIDIA DGX Spark Founders Edition (128 GB, 4 TB) across the EU on one contract and invoice, with manufacturer warranty, for example two systems for a model above 128 GB or four for a team’s shared test environment. NVIDIA AI Enterprise for DGX Spark can go on the same quote. Where a team needs interactive speed for many users, we build AI servers to order with RTX PRO 6000 Server Edition or H200 NVL cards, and we check the rack, power and airflow before we quote.

FAQ