An RTX 5090 costing around RMB 30,000 and equipped with 32GB of VRAM may still be unable to fit the large language model you want to run locally. This illustrates that competition in local AI hardware is no longer merely a contest of GPU compute power; it is about matching memory capacity, bandwidth and model architecture.
What This Is
The article compares an RTX 5090 with a mini PC powered by the AMD Ryzen AI Max+ 395. The RTX 5090 delivers roughly seven times the single-precision compute performance and has a memory bandwidth of up to 1,792GB/s. By contrast, the Ryzen AI Max+ 395 system offers 128GB of unified memory shared by the CPU and GPU, with system memory available to the processor, making it better suited to large models.
However, fitting a model into memory does not mean it will run quickly. Consider a model with 40GB of weights that generates one token at a time, the smallest unit of text produced by a model. The AMD system is constrained by 256GB/s of memory bandwidth and can theoretically read the full set of weights only about six times per second. Longer prompts are further constrained by its lower compute performance: processing a 4,000-token document could take more than 40 seconds.
Mixture-of-Experts models, or MoE, an architecture that activates only some parameters for computation, change this tradeoff. Qwen3-30B-A3B has approximately 30.5 billion total parameters, but activates only around 3.3 billion for each token, allowing it to run on unified-memory devices at lower cost.
Industry View
We believe the article’s most important judgment is that buying the most expensive discrete GPU does not necessarily deliver the best local AI experience. For users who prioritize capacity, power consumption and device size, a unified-memory mini PC may be more practical. For those who prioritize generation speed and long-prompt processing, a discrete GPU with high bandwidth and ample VRAM still has clear advantages.
That said, the claim that “unified memory will win” still lacks sufficient evidence. The usable speed of an MoE model depends not only on its number of active parameters, but also on memory access efficiency, framework optimization, concurrency patterns and thermal performance. In addition, when the article cites version 1017 of DeepSeek V4.1 Flash and estimates its speed using only 16 billion active parameters and 4-bit precision, that theoretical calculation cannot substitute for end-to-end testing on actual hardware.
Impact on Regular People
For enterprise IT: If employees need to process sensitive documents locally, first assess whether the model fits in memory, then evaluate generation speed and long-text performance. Do not equate a GPU’s price directly with productivity.
For individual professionals: Running local models consumes substantial memory and compute resources, and an underpowered configuration can slow down office applications. MoE lowers the barrier to entry, but that does not mean an ordinary PC can match the speed of a cloud-hosted model.
For the consumer market: The promise of running an extremely large model on a single device will continue to drive demand for mini PCs, Mac Studio systems and similar products. Even so, model update speed, quantization support and real-world throughput may be more useful comparison metrics than the parameter counts advertised by vendors.