GKRootWire
AI Google Adds 'Preferred Source' Button to Help Publishers Fight AI Traffic LossesGadgets Linkdaze Launches a Smart Calendar Aimed at Running Your Whole HouseholdSecurity Popular Rust Crate arrayref Hijacked to Spread Infostealer MalwareCloud & Sysadmin GitHub Details Cause of August 17 Outage, Outlines Reliability FixesDev Tools Show HN: 'Huzzah' Proposes a Fresh Take on AI-Assisted CodingCloud & Sysadmin The Weird Science of Cooling Data Centers With UrineAI Google Adds 'Preferred Source' Button to Help Publishers Fight AI Traffic LossesGadgets Linkdaze Launches a Smart Calendar Aimed at Running Your Whole HouseholdSecurity Popular Rust Crate arrayref Hijacked to Spread Infostealer MalwareCloud & Sysadmin GitHub Details Cause of August 17 Outage, Outlines Reliability FixesDev Tools Show HN: 'Huzzah' Proposes a Fresh Take on AI-Assisted CodingCloud & Sysadmin The Weird Science of Cooling Data Centers With Urine
Dev Tools

Pure C GPT Implementation Hits 10M Tokens/Sec on Apple M5

A minimalist GPT clone written in plain C squeezes surprising throughput out of Apple's latest silicon.

A developer has released MicroGPT-C, a lightweight GPT-style language model implementation written entirely in C with no external dependencies. On Apple's new M5 chip, the project reportedly reaches 10 million tokens per second in its benchmark tests, a number that has caught attention on Hacker News despite the project's modest size.

The appeal isn't that this is a production-ready LLM — it's a compact, hackable reference implementation that strips away the layers of Python, CUDA, and framework overhead typically standing between a developer and the raw mechanics of a transformer model. Being pure C also means it can run practically anywhere with a C compiler, including embedded or resource-constrained environments.

Projects like this tend to serve as educational tools and performance testbeds rather than deployment targets, but the throughput numbers highlight just how much headroom modern Apple Silicon offers for hand-tuned, dependency-free code.

Why it matters: Minimal, dependency-free implementations like this matter because they make transformer internals inspectable and portable, useful for teaching, embedded inference, or squeezing performance without a bloated ML stack. It's also a nice benchmark of Apple M5's raw compute potential outside the usual PyTorch/MLX ecosystem.

Sources: Hacker News