Tag
Go introduces experimental SIMD APIs to enable platform-independent SIMD operations, aiming to provide near-asm performance across different CPU architectures.
Google shares open-source vectorized Quicksort code that achieves 10x speedup over C++ std::sort by leveraging SIMD instructions, offering performance-portable sorting across modern CPU architectures.
Zyphra Research releases PUFFER, an incremental fuzzy deduplication system for LLM-scale datasets that runs entirely on CPU with 11-35x speedup over existing methods, open-sourced under Apache 2.0.
A developer created a minimal 700-line C implementation for running the Gemma 4 E2B LLM on CPUs, outperforming llama.cpp in speed.
This draft pull request introduces AVX2 optimizations to speed up large batch size prompt processing for IQ models in llama.cpp, an open-source tool for efficient LLM inference.
FlashAttention-V introduces a blocked FlashAttention optimization for scalable vector architectures, achieving up to 42× speedup in transformer inference for small language models on CPUs and identifying quantization bottlenecks.
A developer reports achieving 92.32% accuracy on the MNIST dataset in 17.6 seconds using a binary XNOR ensemble model on an AMD Ryzen CPU, demonstrating significant speed improvements over traditional floating-point CNNs.
Claude Code CLI has been optimized to reduce CPU usage by 2x at p99 by updating Bun's garbage collector to run only when the process is idle.
Venkat explains that unoptimized CPU work in the hot path can severely impact inference performance, and introduces his PR to mooncake that adds a memory arena for lock-free, allocation-free operations, benefiting vLLM and SGL projects.
An in-depth technical blog post explaining how to efficiently transpose matrices using SIMD instructions on modern x86_64 CPUs, focusing on AVX2 intrinsics like _mm256_shuffle_epi8.
Pull request adds optimized x86 and generic CPU q1_0 dot-product kernels to ggml-cpu, improving quantized LLM inference speed.
Research on optimizing 2D graphics rendering on CPUs using sparse strip techniques to improve performance and reduce memory overhead.
微软官方 BitNet 仓库（bitnet.cpp）的持续更新合集：提供 1-bit/1.58-bit LLM 的高效 CPU/GPU 推理框架，近期新增 BitNet-embedding 0.6B/270M 嵌入模型、I2_S 量化优化内核以及实时多语种 ASR 推理引擎 VibeASR.cpp。