PulseAugur
EN
LIVE 17:53:10

New Runtimes and Benchmarks Boost LLM Inference on Apple Silicon

By PulseAugur Editorial · [3 sources] ·

Researchers have developed new methods for optimizing large language model (LLM) inference on Apple Silicon. The first approach, BaseRT, is a native Metal runtime that achieves higher inference throughput than existing frameworks by optimizing for Metal's execution model and Apple Silicon's unified memory. It supports various model families and quantizations, demonstrating significant performance gains on M3 and M4 Pro devices. The second contribution, Metal-Sci, is a benchmark suite designed for evaluating LLM kernel search on Apple Silicon, featuring ten tasks across different optimization regimes. This benchmark, when used with an evolutionary search harness, has shown substantial in-distribution speedups for models like Claude Opus 4.7, Gemini 3.1 Pro, and GPT 5.5, while also highlighting the importance of out-of-distribution testing to catch silent regressions. AI

IMPACT Optimizations for Apple Silicon could enable more powerful on-device AI applications, reducing reliance on cloud infrastructure.

RANK_REASON The cluster describes novel research papers detailing new software runtimes and benchmarks for LLM inference on specific hardware.

Read on arXiv cs.AI →

AI-generated summary · Google Gemini · from 3 sources. How we write summaries →

New Runtimes and Benchmarks Boost LLM Inference on Apple Silicon

COVERAGE [3]

  1. arXiv cs.AI TIER_1 English(EN) · Prabod Rathnayaka, Fabian Waschkowski, Lukas Wesemann ·

    BaseRT: Best-in-Class LLM Inference on Apple Silicon via Native Metal

    arXiv:2607.00501v1 Announce Type: cross Abstract: We present BaseRT, a native Metal inference runtime for large language models (LLMs) on Apple Silicon, and report the highest inference throughput on this hardware to date. Existing runtimes, including llama.cpp and MLX-based fram…

  2. arXiv cs.AI TIER_1 English(EN) · Lukas Wesemann ·

    BaseRT: Best-in-Class LLM Inference on Apple Silicon via Native Metal

    We present BaseRT, a native Metal inference runtime for large language models (LLMs) on Apple Silicon, and report the highest inference throughput on this hardware to date. Existing runtimes, including llama.cpp and MLX-based frameworks, incur overhead from abstractions not desig…

  3. arXiv cs.AI TIER_1 English(EN) · V\'ictor Gallego ·

    Metal-Sci: A Scientific Compute Benchmark for Evolutionary LLM Kernel Search on Apple Silicon

    arXiv:2605.09708v2 Announce Type: replace-cross Abstract: We present Metal-Sci, a 10-task benchmark of scientific Apple Silicon Metal compute kernels spanning six optimization regimes (stencils, all-pairs in $n$-body problems, multi-field Boltzmann, neighbor-list molecular dynami…

RELATED ENTITIES

RELATED TOPICS