@ggerganov @slaren Where are the other Zen DLLs?

[OPS-NeoRetro]

@ggerganov @slaren Where are the other Zen DLLs?

I found out that llama.cpp and GGML in general don't have microarchitectural optimizations for AMD Zen 1, Zen+, Zen 2 and Zen 3. There are no dedicated DLLs or libggml-cpu for Zen 2 and Zen 3 even when tens of millions of Zen 2 and 3 CPUs are still in service and sold as new parts in new laptops and devices in 2026, including the Steam Deck and Steam Machine. There are many Ryzen 7000 laptop CPUs that still use Zen 2 and Zen 3 CPU cores, like the Ryzen 5 7520U and Ryzen 7 7735H. They require much-needed GGML optimizations to run open-source and open-weight transformer models efficiently on CPU fallbacks.

When app developers start to bundle Llama.cpp and Whisper.cpp instead of TFLite/LiteRT to run transformer models in their apps, many older devices and PCs with weak GPUs and/or NPUs that are still capable of running transformer models will be left behind. Users will blame app developers for slow AI features instead of Llama.cpp or Whisper.cpp's developers.

The Evidence

From my llama-b9442-bin-win-cpu-x64 directory:

• 11 Intel DLLs (2008-2023)
• 1 AMD Zen 4 DLL
• 1 AMD Piledriver DLL
• 0 AMD Zen 2/3 DLLs

From llama-b9442-bin-win-cpu-arm64:

• 1 generic ARM DLL
• Linux ARM64 has 8 optimized libraries. Windows on ARM has 0.

Discussion

• Which CPU/platform without optimizations must GGML/llama.cpp/whisper.cpp optimize for first?
• What CPU/platform are you developing on?
• What performance gains would justify this work?
• Would you sponsor or test?