Tachyon is a new open-source IPC library that hits 56.5 nanosecond round-trip times between processes. That's roughly 13.2 million round-trips per second on a standard Intel i7-12650H. The trick? Skipping the kernel entirely on the hot path and communicating through shared memory with zero-copy semantics.
Created by developer riyaneel, the core is written in C++23 exposing a C ABI, with bindings for Rust, Python, Go, Java, and Node.js. If you're stitching together services in multiple languages and serialization overhead is eating your budget, Tachyon is built for that gap. The project documents every major design choice in Architecture Decision Records, from why it uses memfd_create over shm_open to why there's no serialization contract at all.
The design is uncompromising. It's strictly single-producer single-consumer, which sounds limiting but removes the need for compare-and-swap loops on the hot path. Every structure is padded to 128-byte boundaries, making false sharing between producer and consumer cache lines impossible by construction. The wait strategy spins briefly then falls back to futex on Linux or __ulock_wait on macOS to avoid CPU starvation.
For comparison, ZeroMQ typically lands in the 10-50 microsecond range for round-trips. Eclipse Iceoryx gets closer with its own zero-copy shared memory approach, but it carries the weight of automotive safety-critical compliance and complex pub-sub semantics. Tachyon strips everything down to chase near-physical-limit speed. That minimalism is the point. Doing Less. If you need pub-sub or multi-consumer fanout, this isn't your library. But if you have two processes that need to talk fast with no overhead between them, Tachyon does exactly that and nothing more.