Topic
Everything filed under Performance, newest first.
RSS · JSON · All topics
Training a 4B model to produce 81% faster query plans than Postgres
Leis et al. asked this exact question in 2015. Then, they asked it again 10 years later. Despite an enormous body of research spanning a decade since their original exploration, they found that query optimizers continue to leave much to be desired. I was surprised when I first learned about this. A Postgres database should know everything about the stuff that lives in its tables, no? How hard can it be?
41 min · 9,393 words
Subnormal floating-point numbers are expensive… on Intel processors
Daniel Lemire benchmarks IEEE subnormal floating-point performance across Intel Granite/Emerald Rapids, AMD Zen 5, AWS Graviton 5, and Apple M4 Max, finding ~45–50× slower multiplies on Intel while AMD and Arm stay near full speed.
2 min · 466 words
Performance Improvements in .NET 11
Take a tour through hundreds of performance improvements in .NET 11.
164 min · 37,742 words
Better Vector Search for Long Documents: Chunking Inside Manticore Search
An embedding model reads only the first few hundred tokens of a document and silently drops the rest. Manticore Search now splits long documents for you at INSERT time: add chunk_strategy to the vector column and pick one of five strategies. No ingest pipeline, no splitter library. On our own manual, recall@5 for deep content went from 55% to 83%.
31 min · 7,181 words
The last mile of a long road: faster NumPy in the browser
Notebook.link explains how Emscripten-forge NumPy now links OpenBLAS in WebAssembly, delivering large matmul and linear-algebra speedups for browser scientific computing.
19 min · 4,315 words
You can run git on object storage if you re-make packfiles
Building ObjGit, Tigris explains why Git packfiles fight object storage, how remaking packfiles unlocks workable remote repositories, and what that means for Git servers backed by S3-style buckets.
15 min · 3,408 words
When the fractional part of a float fixes your shader
Bruno Croci debugs a Voronoi Shadertoy stutter that only appeared on a Windows RTX 4070: ANGLE/FXC optimizations dropped `fract` for integer-looking inputs, and swapping a literal to `398.1` or using `x-floor(x)` restored smooth motion.
1 min · 288 words
A 40ms Go GC pause caused by swap
Running swap in production to absorb memory spikes, the author found that Go's garbage collector reads off-heap metadata during a stop-the-world pause, and when that metadata is swapped out the pause jumps from a median of about 51 microseconds to 40ms. Includes eBPF measurements and a note on Go 1.26's Green Tea GC.
3 min · 785 words
Alexandre Lamure walks through first-principles differences when porting a small OpenGL engine to Vulkan—bindless resources, explicit synchronization, and modern Vulkan 1.3 patterns—with pointers to nebula and further reading.
9 min · 2,040 words
Principles for fast Tokio applications
A practical guide to building low-latency, high-throughput applications on the Tokio async runtime, written from experience debugging real production systems. The post covers measuring schedule latency, the fairness-versus-batching trade-off, mutex pitfalls, isolating Tokio workers from other threads, and when multiple runtimes make sense.
1 min · 295 wordsagent-written
I made a build profiler to understand Bun's compile times
Lalit Maganti built buildprof, an open-source Linux tool that records every process spawned during a build and renders them on a shared timeline. He used it to investigate the 5x speed difference between Bun's Zig and Rust builds, tracing the bottleneck to Full LTO in the linker and a downloaded WebKit library that amplified it.
1 min · 318 wordsagent-written
Eddie argues that Pandas forces data practitioners to adopt distributed compute infrastructure long before their data sizes justify it, and that DuckDB and Polars can fill the gap for workloads up to around 100 GB on a single machine. Benchmark results show Polars and DuckDB completing a one-billion-row task in under a minute, while Pandas takes over twelve.
1 min · 268 wordsagent-written
How sparse resources helped my GPU-driven renderer memory usage
A deep dive into using sparse/reserved GPU resources (D3D12-focused) to manage memory for GPU-driven renderer data structures like dynamic and bit arrays.
11 min · 2,502 words
RTK reports huge token savings, but our cost benchmarks disagree
Quesma ran RTK (Rust Token Killer) against Terminal-Bench 2.1 across 1,740 attempts with Claude Code and DeepSeek, and found that compressing terminal output does not reliably reduce cost: Fable saved 3% on a per-pass basis and only because of one anomalous task, while DeepSeek became 7% more expensive.
1 min · 326 wordsagent-written
Tuning a Server for Benchmarking
How to tune a Linux server so benchmarks are repeatable: isolating noise from CPU frequency scaling, interrupts, and background services so small performance wins are actually visible.
5 min · 1,109 words
Introducing chdb Postgres extension: High-performance imports from cloud storage
ClickHouse announces the chdb Postgres extension: fast imports and exports across cloud storage and formats, powered by the embedded ClickHouse engine and usable via COPY-style workflows.
7 min · 1,599 words
Cohere's North Mini Code Megakernel Serving Engine
Today, Cohere presents a serving engine for North Mini Code built around a decode megakernel: BF16 on a single H100, 1.25× - 1.41× faster than vLLM end-to-end. Explore the code behind the serving engine on GitHub. Most LLM serving stacks still treat each forward pass as a sequence of kernels: launch QKV, wait; launch attention, wait; launch the MoE, wait. Each launch is fine on its own. The problem is the waiting in between. At small batch sizes, the GPU spends a surprising fraction of every decode step waiting rather than computing.
24 min · 5,534 words
Introducing CUDA Rust: Two Tracks for Writing GPU Kernels
In September 2026, NVIDIA announced it is leaning into native GPU programming in Rust. CUDA C++ and CUDA Python are mature, enterprise-grade toolchains, and NVIDIA will be growing and maturing CUDA Rust into 2027 and beyond The systems layer of AI spans inference engines, serving infrastructure, drivers, and agent runtimes, and it churns constantly as models and techniques change. More and more of it is written in Rust, which catches whole classes of bugs at compile time…
11 min · 2,528 words
vLLM x AgentX: Optimizing for Real-World Agentic Serving
**TL;DR:** Agentic workloads are becoming a major source of vLLM traffic. Their multi-turn sessions, long contexts, and extensive prefix reuse demand optimizations across the serving stack. This post walks through vLLM's coordinated approach: KV cache management, parallelism and engine optimizations, and methodologies for prefill/decode disaggregation. Measured on AgentX, SemiAnalysis's public agentic benchmark, vLLM achieves up to 130K total tokens per GPU-second on DeepSeek V4 Pro, and an…
17 min · 3,923 words
Serving LLMs on Tenstorrent Hardware: Inside the vLLM TT Plugin
vLLM and Tenstorrent introduce an out-of-tree TT plugin that registers Tenstorrent accelerators as a vLLM platform—covering mesh scheduling, single-process data parallel choices, and an unchanged OpenAI-compatible serving surface.
12 min · 2,863 words