RTX 5090 vs Ryzen AI Max+ 395: running local LLMs

On November 10, 2026, a technical post on Juejin compared two kinds of hardware for running large language models locally. One is NVIDIA's discrete RTX 5090, which carries 32GB of video memory and costs around CNY 30,000. The other is an all-in-one mini PC built on an integrated chipset, such as a machine with the AMD Ryzen AI Max+ 395 (Strix Halo), configurable with up to 128GB of unified memory, a build priced around CNY 20,000. Measured by single-precision FP32 throughput, the discrete card is roughly seven times faster than the integrated chipset. The post also lists the two storage limits that matter most here, RTX 5090 at 1792GB/s of bandwidth and Ryzen AI Max+ 395 unified memory at only 256GB/s.

The writer argues that higher compute does not guarantee you can actually run the model you want, because memory capacity can become the binding constraint first. Take a 70B parameter model quantized to 4 bits per weight, which needs about 35GB (32.6 GiB) of weights alone, more than RTX 5090's 32GB can hold. A partial workaround is placing part of the weights in system RAM, at the cost of speed. The unified memory machine, where the GPU and CPU share one memory pool, has room to spare for such a model, and Apple's M-series chips use the same design. However, unified memory is capped by bandwidth, so token generation could fall back to a few tokens per second. The post adds that the MoE architecture cuts how many parameters must be read per token, easing pressure on bandwidth. No benchmark numbers are included, only this spec-and-principle analysis.