Nvidia has unveiled its highly anticipated AI computing solutions, revolutionizing how professionals run machine learning models and LLMs locally. The newly announced DGX Spark and DGX Station promise unprecedented performance for AI workloads, though with some surprising specification details worth examining before making that significant investment.
Nvidia DGX Spark Mini PC: Impressive Yet Expensive
The DGX Spark (formerly known as Project Digits) has finally arrived as Nvidia’s compact AI desktop solution. Priced at $4,000, this tiny powerhouse offers some compelling specs for running large language models, though not without limitations.
Key specifications include 128GB of unified system memory (non-upgradeable LPDDR5X), similar to Apple Silicon’s approach. However, its memory bandwidth comes in at just 273GB per second—equivalent to Apple’s M4 Pro, not even matching the M4 Max. For comparison, the RTX 4090 delivers nearly four times this bandwidth, albeit with memory limitations.
Perhaps more disappointing is the tensor performance rating of 1,000 AI TOPs, lower than even the RTX 5070. For $4,000, many expected more from this specialized AI machine. Asus has already announced their variant, the Ascent GX10, at a slightly more palatable $3,000.
The Game-Changer: DGX Station Specifications
While the DGX Spark might leave some underwhelmed, the professional-grade DGX Station presents a truly impressive alternative for serious AI work. This desktop-sized machine packs specifications that could transform how data scientists and AI researchers work:
- Blackwell Ultra GPU with up to 288GB of GPU memory
- Staggering 8TB per second of memory bandwidth
- 496GB of CPU memory with 396GB/s bandwidth
- NVLink technology allowing multiple GPUs to connect at 900GB/s
- Nvidia Connects networking at up to 800Gb/s
- An incredible 20 petaFLOPs of AI computing power
This workstation is clearly engineered specifically for AI workflows, though pricing hasn’t been announced yet. Given the DGX Spark’s pricing, expect the DGX Station to come with a substantial price tag targeting enterprise customers and serious AI professionals.
Memory: The Critical Factor for LLM Performance
When running large language models locally, two factors determine performance: memory bandwidth and total available memory. Consumer GPUs like the RTX 4090 and 5090 offer incredible bandwidth (up to 1,700GB/s for the 5090) but are limited to 32GB of VRAM.
This creates a fundamental tradeoff: faster memory versus more memory. When models exceed available GPU memory, processing spills over to CPU memory, dramatically reducing performance. The DGX Station’s massive 288GB of GPU memory with 8TB/s bandwidth effectively eliminates this bottleneck.
Clustering Capabilities: Scaling AI Processing
Both the DGX Spark and DGX Station support clustering, allowing users to connect multiple units for increased performance. However, the connection between machines becomes the new bottleneck.
The DGX Station uses NVLink technology providing 900GB/s bandwidth between units—still impressive but notably lower than its internal 8TB/s memory bandwidth. For networking between systems, Nvidia introduces its Connects technology offering up to 800Gb/s, far surpassing traditional 10GbE or 25GbE connections.
Floating Point 4: A New Approach to AI Calculation
Interestingly, Nvidia CEO Jensen Huang mentioned the performance metrics were based on floating point 4 calculations rather than the typical FP16 or FP32. This suggests a specialized approach to AI computation that may affect how models perform compared to traditional quantization methods like INT4.
As AI capabilities continue to accelerate development across industries, Nvidia’s new offerings provide specialized hardware solutions for different tiers of AI professionals. While the DGX Spark may appeal to individual developers with its compact form factor, the DGX Station represents a true paradigm shift for organizations requiring maximum AI performance without the complexity of server racks. For those looking to seriously integrate AI into their workflow in 2025, these options—particularly the DGX Station—offer compelling, if expensive, solutions to the computational challenges of modern AI.