The NVIDIA DGX Spark and the NVIDIA GeForce RTX 4090 are two distinct computing systems designed for different purposes, leading to significant differences in their architectures.
CPU Architecture
- DGX Spark: It is powered by the NVIDIA GB10 SoC, which features a 20-core Armv9 processor. This includes ten Cortex-X925 cores and ten Cortex-A725 cores, designed for high-performance AI processing and general computing tasks[1][4].- RTX 4090: This GPU does not have a dedicated CPU; instead, it is a graphics processing unit (GPU) designed to work with a host CPU. It is based on the Ada Lovelace architecture, specifically the AD102 GPU, which is optimized for gaming and graphical workloads[2][5].
GPU Architecture
- DGX Spark: The GPU in the DGX Spark is based on the Blackwell architecture, which is part of the NVIDIA Grace Blackwell Superchip. It is optimized for AI workloads, providing up to 1,000 TOPS of AI performance[1][7].- RTX 4090: The RTX 4090 features the Ada Lovelace architecture, which includes 16,384 CUDA cores, 128 Ray Tracing cores, and 512 Tensor cores. This architecture is primarily focused on gaming, ray tracing, and high-performance computing tasks[2][8].
Memory and Bandwidth
- DGX Spark: It uses 128 GB of 256-bit LPDDR5x memory, offering a memory bandwidth of 273 GB/s. The system also employs NVIDIA NVLink-C2C interconnect technology for enhanced CPU-GPU communication, providing five times the bandwidth of PCIe 5.0[1][4].- RTX 4090: The RTX 4090 comes with 24 GB of GDDR6X memory and a memory bandwidth of 1008 GB/s. It uses a PCIe 4.0 x16 interface for connection to the host system[2][5].
Tensor Cores and AI Performance
- DGX Spark: Equipped with 5th Generation Tensor cores, the DGX Spark is optimized for AI tasks, offering up to 1,000 TOPS of AI performance. This makes it suitable for training and inference of large AI models[1][7].- RTX 4090: Although primarily a gaming GPU, the RTX 4090 includes 4th Generation Tensor cores, which enhance its capabilities for AI and machine learning tasks. However, its AI performance is not as high as specialized AI systems like the DGX Spark[2][8].
Networking and Connectivity
- DGX Spark: It features a 10GbE RJ45 port via ConnectX-7 Smart NIC, WiFi 7, Bluetooth 5.3, and four USB4 ports. The ConnectX-7 allows for networking capabilities that can connect two DGX Spark systems for larger AI models[1][4].- RTX 4090: As a GPU, it does not have built-in networking capabilities but relies on the host system's connectivity options. It connects via a PCIe 4.0 interface[2][5].
In summary, the DGX Spark is designed as a compact AI supercomputer with a focus on AI processing and research, while the RTX 4090 is a high-performance GPU optimized for gaming and graphical workloads.
Citations:
[1] https://www.cnx-software.com/2025/03/19/nvidia-dgx-spark-a-desktop-ai-supercomputer-powered-by-nvidia-gb10-20-core-armv9-soc-with-1000-tops-of-ai-performance/
[2] https://www.itcreations.com/nvidia-gpu/nvidia-geforce-rtx-4090-gpu
[3] https://www.fibermall.com/blog/nvidia-a100-vs-4090.htm
[4] https://www.pcmag.com/news/what-is-nvidias-dgx-station-a-new-specialized-desktop-line-for-ai-work
[5] https://www.techpowerup.com/review/nvidia-geforce-rtx-4090-founders-edition/2.html
[6] https://beebom.com/nvidia-rtx-5070-vs-rtx-4090-comparison/
[7] https://www.constellationr.com/blog-news/insights/nvidia-launches-dgx-spark-dgx-station-personal-ai-supercomputers
[8] https://coinpoet.com/ml/learn/gpu/nvidia-geforce-rtx-4090
[9] https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf