DGX Spark with Qwen3: An NVIDIA Test That Doesn't Tell the Whole Story

DGX Spark with Qwen3: An NVIDIA Test That Doesn't Tell the Whole Story

As large language models (LLMs) grow more powerful and more demanding, running them runs into a acute memory shortage. Many enthusiasts and even professionals face the same situation: a model with tens of billions of parameters simply does not fit into the limited video memory (VRAM) of modern graphics processors. That forces a compromise.

Either you offload part of the model to the central processor (CPU), which slows it down catastrophically, or you turn to expensive cloud services, which can also raise data privacy concerns. To address this pressing problem, NVIDIA introduced the DGX Spark system (also known as GB10), equipped with 128 GB of unified memory. Its stated price is 400,000 to 500,000 rubles.

But a two-week deep test showed that the system's results with the Qwen3 model were mixed, which raises questions about the real value of the product.

Anyone who works with LLMs knows the problem DGX Spark is meant to solve. A typical case: trying to run a 32-billion-parameter model on an RTX 4090-class graphics card with 24 GB of VRAM. A memory shortfall is unavoidable.

Offloading to the CPU lets you run the model, but performance drops to an unacceptable level. Cloud services, for their part, demand significant spending and also put the security and privacy of the processed data in question, since it is sent to third-party servers. Against this background, NVIDIA's offer of DGX Spark with 128 GB of unified memory looks promising.

Unlike the traditional split between CPU and GPU memory, unified memory lets both processors work on the same data without copying it, which in theory should speed up processing.

A deep dive into testing DGX Spark with Qwen3 revealed a number of nuances. The benchmarks were extensive: they compared different quantization formats (a method that shrinks a model and its memory requirements at the cost of some accuracy), tested different context sizes (the amount of information the model processes at once) and compared performance with more familiar GPU setups. The results were far from clear-cut.

In some scenarios DGX Spark did show an advantage, particularly when it had to handle volumes of data that did not fit into the VRAM of standard graphics cards. In other cases, especially under heavy load or with certain types of models, the system did not deliver the expected speedup. Moreover, DGX Spark's performance sometimes fell short of well-optimized multi-GPU setups, or required turning to expensive cloud resources, which negates the device's main advantage: local data processing.

It is worth noting that in some tests, when the model did not fit entirely into unified memory, the system switched to the CPU automatically, causing a significant slowdown comparable to ordinary offloading.

Such mixed results call for careful analysis. DGX Spark's cost-effectiveness, at a price of half a million rubles, is questionable, especially since in a number of scenarios it offers no tangible advantage over cheaper or more conventional solutions. The system's architectural limits, which show up under certain types of workloads, make it a niche product rather than a universal one. NVIDIA's marketing materials likely emphasize peak performance figures and scenarios where 128 GB of memory is truly the deciding factor, while saying nothing about situations where this architecture may prove inefficient or even a losing proposition. This means prospective buyers need to weigh their own tasks carefully against DGX Spark's real capabilities rather than rely solely on promotional claims.

In conclusion, DGX Spark with 128 GB of unified memory is an interesting but imperfect step in hardware for working with large language models. It can pay off for a narrow range of tasks where memory capacity is critical and other solutions simply cannot cope. For most users, however, 128 GB of unified memory is not a "silver bullet" that solves every problem. We have to admit that for maximum performance and economic sense, optimized multi-GPU setups or even hybrid approaches often remain the better choice. Careful testing and an understanding of DGX Spark's architectural features are the key to deciding whether the device is really worth its considerable price for your specific needs.