NVIDIA's Project DIGITS, unveiled at CES 2025 as a $4,699 personal AI supercomputer, promised to democratize local AI development with a GB10 Grace Blackwell chip, 128 GB of unified memory, and 1 PFLOP (petaflop, or one quadrillion floating-point operations per second) of compute power. However, early adopters report that the DGX Spark hardware fails to deliver on nearly every marquee specification, raising questions about whether NVIDIA shipped an unfinished platform and expected customers to debug it.

What Are the Real Performance Issues With NVIDIA's DGX Spark?

The DGX Spark's headline specifications crumble under real-world testing. The dual 200 gigabit-per-second networking ports, designed to link multiple units for running massive language models, deliver only about 13 gigabits per second in practice, an order of magnitude below the promise. The root cause is a power-budget bug in the firmware that throttles the network interface even when NVIDIA's own power supply is connected. The system detects insufficient power on the PCIe (PCI Express) slot and automatically reduces performance, a fundamental architectural limitation that firmware updates alone cannot fix.

The networking problem runs deeper than a simple bug. To reach the advertised 200 gigabits per second, users must manually bind both network connections and configure multi-host PCIe aggregation, a setup most customers never attempt. The GB10 system-on-chip can physically provide only about 100 gigabits per second to a single device through PCIe Gen5 x4 connectivity. Daisy-chaining three units cuts bandwidth in half again, forcing users to purchase expensive network switches.

The memory bandwidth constraint presents another hard ceiling that no software update can overcome. The 128 GB of unified memory is shared between the CPU and GPU through LPDDR5X memory running at 273 gigabytes per second. For token generation, the process of producing text one word at a time, this bandwidth becomes the bottleneck. On a 20-billion-parameter model, the Spark achieves only 49.7 tokens per second during generation, while a single RTX 5090 graphics card hits 205 tokens per second, and an Apple Mac Studio M4 Max delivers roughly double the bandwidth at a similar price.

Why Is NVIDIA's NVFP4 Software Feature Broken?

The 1 PFLOP headline number depends entirely on NVFP4, NVIDIA's proprietary 4-bit floating-point format. This feature is the most visibly broken component of the shipping software stack. One customer who invested approximately $38,000 across nine DGX Spark units publicly demanded a roadmap after discovering the promised software was not in a usable state.

The software failures span multiple critical areas:

• Qwen3.5 NVFP4 Models: Crash with CUDA (Compute Unified Device Architecture, NVIDIA's parallel computing platform) illegal instruction errors on ARM64 GB10 processors.
• Nemotron-3-Nano Models: Trigger CUDA errors during graph capture for batch sizes larger than one, preventing efficient batch processing.
• Mixture-of-Experts Models: Hit misaligned-address errors because the workspace buffer does not meet stricter alignment requirements.
• Build System Gaps: For an extended period, the SM121 architecture guards were missing entirely from vLLM's build system, meaning NVFP4, CUTLASS, and MLA kernels were silently skipped at compile time, forcing users to run slower fallback paths without knowing it.

NVIDIA has had a year since the DGX Spark's launch to resolve these issues, yet they persist in production environments. This is not typical early-adopter friction; it represents a fundamental failure to deliver promised functionality.

How to Evaluate Whether the DGX Spark Makes Financial Sense

• Price-to-Performance Comparison: The DGX Spark costs $4,699, while a Framework Desktop with AMD Strix Halo (offering 128 GB unified memory and comparable token generation speed) costs $2,348, less than half the price. A used cluster of three RTX 3090 graphics cards can be assembled for under $2,000 and triples the decode speed for models that fit in memory.
• Memory Bandwidth Reality: The Spark's 273 gigabytes-per-second bandwidth is shared between CPU and GPU, creating a hard architectural ceiling. The Mac Studio M4 Max at $3,999 delivers double the memory bandwidth, making it faster for token generation despite costing less.
• Software Maturity Risk: Purchasing a DGX Spark today means accepting the role of a beta tester on a platform where the three top-line marketing claims do not hold up under scrutiny. Fixes that have arrived, such as hot-plug support and Wi-Fi detection, are basic features, not solutions to architectural bugs like the 27-watt power throttle.

The DGX Spark's only defensible niche is CUDA-on-ARM development with fast model prefill, the initial processing phase that loads the model into memory. However, this advantage requires waiting for software fixes that may never arrive. Even NVIDIA co-founder John Carmack benchmarked his own DGX Spark and found it drawing only about 100 watts of system power, far below the rated 240 watts, with correspondingly reduced performance.

What Hardware and Software Issues Are Affecting Early Adopters?

Beyond the three core failures, documented issues include faulty power supplies, units that become unusable after firmware updates, and complete network stack failures on arrival. These are not isolated incidents but recurring problems affecting multiple customers across forums and support channels.

The broader pattern suggests NVIDIA prioritized marketing claims over engineering rigor. The company doubled the prices of its RTX 5090 graphics cards from approximately $2,500 to $5,000 in early 2026 with no corresponding performance improvement, a move characterized as price gouging rather than value-based pricing. Combined with the DGX Spark's unfinished state, this pricing strategy signals a shift in NVIDIA's priorities away from customer satisfaction and toward stock price appreciation.

For customers considering local AI development, the landscape has shifted. AMD's Strix Halo platform delivers comparable inference speed at typically less than half the price and consumes significantly less power. Apple's Mac hardware offers much faster unified memory bandwidth and is experiencing long wait times due to surging demand from AI developers. NVIDIA's historical dominance in AI hardware, built on CUDA's software ecosystem and reliable driver support, is being challenged by competitors offering better value and more mature software stacks.

The DGX Spark represents a fascinating concept that failed in execution. Decentralized AI development, the vision of running large language models locally without cloud dependency, deserves better than a $4,699 platform where none of the marquee specifications deliver as promised.