RTX 5090 vs. Ryzen AI Max+ 395 for local AI models
An analysis compares NVIDIA’s RTX 5090 with an AMD Ryzen AI Max+ 395 Strix Halo mini-PC for running local AI models. It lists 32GB of VRAM, 1,792GB/s bandwidth and 104.8 TFLOPS of FP32 performance for the RTX 5090, versus 128GB of shared memory, 256GB/s bandwidth and 14.8 TFLOPS for the AMD system. The analysis says the AMD system’s larger unified-memory capacity can accommodate models that do not fit in the RTX 5090’s VRAM, while the RTX 5090’s higher bandwidth may favor workloads that repeatedly read weights, including MoE architectures that activate only part of their parameters. Its token-rate comparisons are theoretical bandwidth-based estimates rather than end-to-end benchmarks.
The comparison shows that a higher-priced local-AI system is not automatically the better choice: memory capacity, bandwidth, prompt processing and model architecture can favor different hardware.
很多人可能不知道,你买的那张好几万的 RTX 5090,可能跑不了你想跑的本地 AI 模型,我可没跟你瞎说。
假设你现在在家运行本地 AI 模型,我给你两种硬件选择。
- 一种是独立显卡,比如英伟达的 RTX 5090,带有 32GB 显存,售价 3 万元人民币左右。
- 另一种是采用板载芯片组(CPU + 板载显卡 + 板载内存)的家用电脑,比如搭载 AMD Ryzen AI Max+ 395(Strix Halo)的迷你电脑,可以配到 128GB 内存,这个配置售价 2 万元人民币左右。
如果是你,你会选哪一种?
大部分人可能不假思索,就会选择独立显卡,毕竟独立显卡的算力远大于板载芯片组,当然,还有另外一个原因,独立显卡贵!!!
#NVIDIA RTX 5090#AMD Ryzen AI Max+ 395#AMD Strix Halo#Qwen3-30B-A3B#DeepSeek V4.1 Flash#Apple M3 Ultra