AMD Boosting AI/LLM Performance For Radeon iGPUs As Much As 18~23% With Linux 7.4

Written by Michael Larabel in Display Drivers on 29 September 2026 at 02:57 PM EDT. Page 3 of 4. 23 Comments.
Lemonade PerfOpt Strix Point

Next up was looking at the impact on Strix Point with the AMD Ryzen AI 9 365 running within the ASUS Zenbook S16 with 16GB of LPDDR5-7500 memory.

Lemonade benchmark with settings of Backend: Vulkan, Scenario: code-debug, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: Vulkan, Scenario: code-debug, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: Vulkan, Scenario: code-debug, Model: Qwen3.5-4B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: Vulkan, Scenario: code-short, Model: Qwen3.5-4B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-debug, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-debug, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-short, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-short, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-debug, Model: Qwen3.5-4B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-debug, Model: Qwen3.5-4B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-debug, Model: MiniCPM4-8B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-debug, Model: MiniCPM4-8B-GGUF. AMD PerfOpt was the fastest.

Well, this is much more interesting... With the smaller memory carve out, Lemonade with Llama.cpp was enjoying much greater gains out of AMD PerfOpt.

Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-short, Model: MiniCPM4-8B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: code-short, Model: MiniCPM4-8B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: Vulkan, Scenario: code-explain, Model: MiniCPM4-8B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: Vulkan, Scenario: code-explain, Model: MiniCPM4-8B-GGUF. AMD PerfOpt was the fastest.

Across many different benchmarks conducted with Lemonade using both the AMD ROCm and Vulkan back-ends of Llama.cpp, there were gains as high as 23% from this new feature coming in Linux 7.4.

Lemonade benchmark with settings of Backend: Vulkan, Scenario: chat-long-output, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: Vulkan, Scenario: chat-long-output, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: chat-long-output, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: chat-long-output, Model: Qwen3-14B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: Vulkan, Scenario: code-short, Model: DeepSeek-Qwen3-8B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: Vulkan, Scenario: code-short, Model: DeepSeek-Qwen3-8B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: chat-long-output, Model: Qwen3.5-4B-GGUF. AMD PerfOpt was the fastest.
Lemonade benchmark with settings of Backend: AMD ROCm, Scenario: chat-long-output, Model: MiniCPM4-8B-GGUF. AMD PerfOpt was the fastest.

Some very nice improvements for throughput and lower latency with AMD PerfOpt on Strix Point with the Radeon 890M integrated graphics. Linux 7.4 just got a whole lot more exciting.

Related Articles