AMDのドライバ更新でAI処理能力が大幅強化され109BのLlama 4 Scoutをローカル実行可能に – GIGAZINE https://www. yayafa.com/2774434/
agents llama meta
| Source: Mastodon | Original article
AMD has rolled out a new driver update for its Ryzen AI Al Max+ 395 accelerator that dramatically expands the chip’s on‑device inference capacity. The patch lifts the practical limit from the 70 billion‑parameter Llama models the hardware already handled to the 109 billion‑parameter Llama 4 Scout, allowing the full model to run locally on a consumer‑grade PC without cloud assistance.
The boost comes from a combination of ROCm‑based kernel optimisations, tighter integration with AMD’s OpenCL‑AI extensions and a revised memory‑management scheme that squeezes more of the 32 GB HBM2e pool into active tensors. Early benchmarks posted by GIGAZINE show a 2.3× speedup on token generation compared with the previous driver, and a 5.4× improvement in image‑generation pipelines that rely on the same compute blocks.
Why it matters is threefold. First, it narrows the performance gap that has long favored Nvidia’s CUDA ecosystem for large‑scale language models, giving developers on AMD platforms a viable alternative for privacy‑sensitive workloads. Second, the ability to run a 109 B model on a desktop lowers the barrier for Nordic startups and research labs that lack the budget for expensive cloud GPU instances. Third, the update signals AMD’s commitment to the AI‑first roadmap it outlined at CES 2025, suggesting that further driver releases could push the ceiling toward the 1.28 trillion‑parameter range hinted at in internal roadmaps.
What to watch next are the upcoming ROCm 6.2 release, which promises broader support for PyTorch and TensorFlow, and AMD’s announced Ryzen AI 500 series, slated for Q4 2026. Industry observers will also be tracking real‑world benchmark suites from the Nordic AI community to see whether the driver gains translate into measurable productivity gains for local LLM deployment.
Sources
Back to AIPULSEN