Pushing CSV to the Limits: SIMD, Zero Allocation, and Branchless Parsing

CSV looks simple, but it is one of the most punishing workloads for modern CPUs. Quoting rules, variable-length fields, and highly unpredictable control flow turn parsing into a perfect storm of branch mispredictions and cache misses.

In this talk I’ll share the journey of building a SIMD-accelerated, zero-allocation CSV parser and emitter. We’ll explore how to scan for delimiters in parallel, how to restructure parsers to be branch-light and cache-friendly, and why iterating over fields instead of records gives API users more control while enabling higher throughput. Along the way, I’ll present real benchmarks, design trade-offs, and the CPU-level reasoning behind them.

We’ll also briefly look at how the same techniques apply to other formats such as JSON, and how—somewhat to my surprise—Zig’s build system and C interoperability made it possible to ship a C ABI with almost the same performance as the native implementation.

The goal isn’t just to build a fast CSV library, but to develop a practical mental model for pushing any text-heavy system closer to the limits of the hardware.

About the speaker...

Peyman Mortazavi is a software engineer with a focus on systems, performance, and developer tools. He enjoys building software with taste — where technical rigor meets thoughtful design — and spends much of his time exploring low-level performance, parsing, and language tooling. When he’s not working on code, he plays the piano, performs in theater, and occasionally knits.

Website: https://peymanmo.com/