---
title: "Discussion on \"Why DDR5 Bandwidth Kills Dual-LLM Inference on APUs (Benchmarks Inside)\" | Hashnode"
description: "Discussion on \"Why DDR5 Bandwidth Kills Dual-LLM Inference on APUs (Benchmarks Inside)\". I benchmarked every dual-model Ollama combo on a 96GB DDR5 APU. Shared memory bandwidth kills the idea. Real numbers, real commands, real conclusions."
url: "https://hashnode.com/posts/why-ddr5-bandwidth-kills-dual-llm-inference-on-apus/6a2405f78b61ee49bf4b147e"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Why DDR5 Bandwidth Kills Dual-LLM Inference on APUs (Benchmarks Inside)](https://joshgreen.hashnode.dev/why-ddr5-bandwidth-kills-dual-llm-inference-on-apus?utm_source=hashnode&utm_medium=feed)

I benchmarked every dual-model Ollama combo on a 96GB DDR5 APU. Shared memory bandwidth kills the idea. Real numbers, real commands, real conclusions.

Jun 6 · 8 min read · joshgreen.hashnode.dev  
Updated Jul 21  
Read the article

-   [#llm](https://hashnode.com/tag/llm)
-   [#self-hosted](https://hashnode.com/tag/self-hosted)
-   [#mini-pc](https://hashnode.com/tag/mini-pc)

## Written by

[Josh Green](https://hashnode.com/@joshgreen)

@joshgreen · Budapest, Hungary

  
[View profile](https://hashnode.com/@joshgreen)

Lead developer at ModelDirectory.org

I frequently automate problems that should not be automated. I then often brag about it.

## Responses

Markdown supported · ⌘↵ to post

No responses yet.