The NVIDIA DGX Spark and the NVIDIA RTX 4090 are both high-performance computing systems, but they serve different purposes and have distinct architectures. Here's a comparison of their memory bandwidth:
NVIDIA DGX Spark
- Memory Type: The DGX Spark uses 128 GB of unified LPDDR5x memory.
- Memory Bandwidth: It offers a memory bandwidth of 273 GB/s, which is facilitated by a 256-bit memory interface[1][4][6].
- Architecture: The system is powered by the NVIDIA GB10 SoC, which includes a Blackwell GPU and supports up to 1,000 TOPS of AI performance. The CPU+GPU-coherent memory model uses NVIDIA NVLink-C2C interconnect technology, providing significantly higher bandwidth compared to traditional PCIe interfaces[1][4].
NVIDIA RTX 4090
- Memory Type: The RTX 4090 features 24 GB of GDDR6X memory.
- Memory Bandwidth: It achieves a memory bandwidth of 1,008 GB/s through a 384-bit memory bus[2][5][8].
- Architecture: The RTX 4090 is based on the Ada Lovelace architecture, which includes 16,384 CUDA cores, 512 Tensor cores, and 128 Ray Tracing cores. It is primarily designed for gaming and high-performance computing tasks like 3D rendering and AI workloads[5][8].
Comparison
In terms of memory bandwidth, the RTX 4090 significantly outperforms the DGX Spark, offering nearly four times more bandwidth (1,008 GB/s vs. 273 GB/s). However, the DGX Spark is optimized for AI workloads with its high TOPS performance and specialized architecture, making it more suitable for tasks like training and inference of large AI models. The RTX 4090, on the other hand, is geared towards a broader range of applications, including gaming and general-purpose computing.
While the DGX Spark's memory bandwidth is lower, its unique architecture and interconnect technology provide advantages in AI-specific tasks. The RTX 4090's higher memory bandwidth makes it more versatile for a variety of high-bandwidth applications.
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://vast.ai/article/nvidia-rtx-4090-vs-5090
[3] https://beebom.com/nvidia-rtx-5070-vs-rtx-4090-comparison/
[4] https://www.pcmag.com/news/what-is-nvidias-dgx-station-a-new-specialized-desktop-line-for-ai-work
[5] https://www.itcreations.com/nvidia-gpu/nvidia-geforce-rtx-4090-gpu
[6] https://www.reddit.com/r/LocalLLaMA/comments/1jef1dd/dgx_spark_previously_digits_has_273gbs_memory/
[7] https://www.youtube.com/watch?v=krBh0Von-2A
[8] https://coinpoet.com/ml/learn/gpu/nvidia-geforce-rtx-4090
[9] https://chipsandcheese.com/p/microbenchmarking-nvidias-rtx-4090
[10] https://www.reddit.com/r/LocalLLaMA/comments/14319ra/rtx_40906000_vs_m2_max_with_96gb_unified_memory/
