Strix Halo
A highly integrated APU platform introduced by AMD, featuring large memory capacity and a unified memory architecture. It is commonly used in mini PCs and small workstations and can support local inference of large language models.
Timeline (last 90 days)
Doug Ware tested Alibaba's open-source model Quen 3.8 Flash Next (a 125-billion-parameter MoE model that activates 6 billion parameters per token) within two weeks on a Strix Halo mini PC with 128GB of memory and an NVIDIA CUDA workstation.
On Strix Halo, GUFO's prefill is about 1400 TPS and output is about 40 TPS.
On Strix Halo, the model's cold-start prefill reached several hundred tokens per second and output held steady at just over 40 TPS; on the CUDA workstation, prefill reached several thousand tokens per second and output sustained between 100 and 2100 TPS.
On Strix Halo with 4-bit quantization (Q4K XL), 42% of tool calls failed because the model wrote an extra `arguments` parameter into the tool definition.
Strata 0.2.3 improves tool-calling speed by about 19% on Strix Halo.
In TDP-constrained form factors such as handhelds, choosing a SKU with fewer CPU cores but a full GPU specification allows more thermal headroom to be dedicated to graphics rendering.
Strix Halo is a high-performance mobile platform launched by AMD in early 2025, with the official product line designated as Ryzen AI Max.
The GPD Win Max 3 handheld is equipped with an AMD Ryzen AI Max+ 395 or 388 (Strix Halo platform), featuring a 9-inch screen, up to 128GB of LPDDR5 memory, a USB 4 interface, and up to 4TB of storage.
By introducing a Vulkan backend to replace the traditional ROCm path, Unsloth brings approximately 20% performance improvement to APUs/iGPUs such as Strix Halo and Strix Point, with prompt processing speed on the Strix Halo platform improving by 23% and generation speed improving by 8%.
All Facts (9)
On Strix Halo with 4-bit quantization (Q4K XL), 42% of tool calls failed because the model wrote an extra `arguments` parameter into the tool definition.
50%UnverifiedOn Strix Halo, GUFO's prefill is about 1400 TPS and output is about 40 TPS.
50%UnverifiedOn Strix Halo, the model's cold-start prefill reached several hundred tokens per second and output held steady at just over 40 TPS; on the CUDA workstation, prefill reached several thousand tokens per second and output sustained between 100 and 2100 TPS.
50%UnverifiedStrata 0.2.3 improves tool-calling speed by about 19% on Strix Halo.
50%UnverifiedIn TDP-constrained form factors such as handhelds, choosing a SKU with fewer CPU cores but a full GPU specification allows more thermal headroom to be dedicated to graphics rendering.
50%UnverifiedStrix Halo is a high-performance mobile platform launched by AMD in early 2025, with the official product line designated as Ryzen AI Max.
50%UnverifiedThe GPD Win Max 3 handheld is equipped with an AMD Ryzen AI Max+ 395 or 388 (Strix Halo platform), featuring a 9-inch screen, up to 128GB of LPDDR5 memory, a USB 4 interface, and up to 4TB of storage.
50%UnverifiedBy introducing a Vulkan backend to replace the traditional ROCm path, Unsloth brings approximately 20% performance improvement to APUs/iGPUs such as Strix Halo and Strix Point, with prompt processing speed on the Strix Halo platform improving by 23% and generation speed improving by 8%.
50%UnverifiedDoug Ware tested Alibaba's open-source model Quen 3.8 Flash Next (a 125-billion-parameter MoE model that activates 6 billion parameters per token) within two weeks on a Strix Halo mini PC with 128GB of memory and an NVIDIA CUDA workstation.
50%