performance-optimization

The article shares a set of prompt techniques for performance optimization with Claude: rewrite code using data-oriented design (DOD), flatten hash tables into arrays, replace pointers with array indices, convert AoS to SoA, and even delete all caches entirely — to truly squeeze hardware performance instead of using caches to mask slow designs.

A GitHub repo shared by a user shows a notable inference speedup for running Qwen3.8 27B locally on the AMD Radeon RX 7900 XTX GPU.

TryMirai open-sourced uzu, an Apple Silicon inference engine, featuring Weaver + DFlash speculative decoding that makes a 9B LLM (Qwen3.5) 4x faster locally and 2.6x faster than standard speculative decoding.

@NVIDIARTXSpark shares three techniques to accelerate Mixture-of-Experts (MoE) inference on the NVIDIA DGX Spark, covering optimized kernels, FP8 quantization, and multi-token prediction to boost tokens per second.

ParadeDB details how segment partitioning (based on the Spooky LSM compaction paper) and other optimizations dramatically sped up COUNT queries and filtered, sorted searches—up to 66x faster than before—on Postgres-based search workloads.

TorchTitan 发布重大性能与架构更新，新增对 DeepSeek-V4、Qwen3.5、Kimi K3、Muse Spark Open 等 SOTA 模型的完整 5D 并行支持，并推出统一后训练与智能体工作负载的 TitanRL 框架，FlexShard 与 torch_checkpointing 大幅提升大规模训练效率。

A Reddit user created an open-source Oblivion Remastered performance mod with the help of Anthropic's Claude AI, delivering 40+ engine-level optimizations that boost FPS by over 40% and greatly reduce stuttering without any visual trade-offs.

gpu-unlock 1.0.0 发布候选版是一个 Linux 内核模块，通过在 NVIDIA 驱动绑定前解锁 RTX 3090 的 BF16 全速档位，免费实现约 15% 的 prompt processing 与 5% 的 token generation 提速，无需修改驱动或 UEFI。

Part 1 of Amazon's tutorial series on profiling TorchNeuron workloads on AWS Trainium, covering how to configure torch.profiler with the Neuron backend, read key_averages() output, and trace operations through Neuron Explorer.

DNSimple 详细介绍了他们如何通过纯 BEAM 二进制处理技巧重写开源 DNS 库 dns_erlang 的热路径，在无需 NIF 或 C 代码的情况下实现每次查询 CPU 降低 12%、每秒查询量提升 14%。

A 2014 blog post by L. Pereira exploring increasingly optimized ways to convert integers to strings in C, comparing naive division-based approaches, an in-place reverse-free variant used in Lwan, and Facebook's pair-of-digits lookup table trick that roughly halves the number of divisions for a ~31% speedup.

介绍 SentryWard：一个开源的 GPU kernel 形式化验证工具，可证明同步安全、无数据竞争与资源生命周期正确性，并通过系统性移除冗余同步使 NVIDIA B200 上的 kernel 最快提速 1.4 倍。

Niles Salter’s talk introduces how to use SIMD data parallelism and bit manipulation primitives (CTZ, TZCNT, etc.) to accelerate a Zig tokenizer, achieving a throughput of 820 MB/s, and explores the potential for further speedups with AVX-512.

curl 8.23.0 引入多项性能优化，包括 base64 编解码、URL 编解码、日期和数字解析提速，以及约 40% 的结构体内存缩减和更少的内存分配。

一篇工程博客分享了如何通过将 Unreal 编辑器着色器哈希从缓慢的 SHA1 替换为引擎内已有的 xxHash，从而加速编辑器启动与资产处理。文章强调应避开已被高度优化的代码路径，转而寻找被忽视的低垂果实。

strata-mlx update: interrupted long conversations no longer get reread — time to first character for a 10K-token session drops from 19.7s to 0.26s (M4 Max), and memory management on 32 GB Macs is optimized to avoid swap. The project is an Apple Silicon MLX inference engine for running 125B MoE models like Qwen3.8-Flash-Next, with throughput significantly better than llama.cpp and LM Studio.

rdna35-perf branch of strix-llama.cpp 为 Strix Halo (Radeon 8060S) 带来新的 RDNA3.5 flash attention 与 GEMM 内核优化，prompt processing 比 Vulkan 快 21–54%，在 128k 长上下文下比原版 ROCm 最快达 2.7 倍，且输出质量与原版 ROCm 基本一致。

Developer helped speed up ROCmFPX, a llama.cpp fork, on AMD Strix Halo (Ryzen AI Max+ 395), achieving major prefill and decode speedups with 16 merged PRs and improved quantization accuracy.

提醒用户尝试 llama.cpp 中新合并的概率式 MTP（多 token 预测）功能，在散文生成场景下解码速度提升约 14%，最优 draft-n-max / draft-p-min 参数与贪心采样基本一致。

NumPy 2.5 introduces a vectorized implementation of np.searchsorted that achieves up to 25× speedup using batching and cache-friendly techniques, benefiting the broader Python scientific ecosystem including SciPy and scikit-learn.