C++26 Shipped a SIMD Library Nobody Asked For

TL;DR

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

C++26 introduces std::simd, a new library for portable SIMD programming. However, experts argue it is slower and less capable than compiler auto-vectorization and existing libraries like Highway. The development raises questions about its practical value.

The C++26 standard has officially included std::simd, a library intended to provide a portable, high-level abstraction for SIMD programming across different architectures. Despite its formal inclusion, experts and benchmarks indicate that std::simd performs poorly compared to compiler auto-vectorization and alternative libraries, raising questions about its practical utility.

Std::simd was developed over nearly a decade, originating from Matthias Kretz’s Vc library, with the goal of allowing developers to write SIMD code once and compile it for various architectures without resorting to intrinsics. The proposal was finalized and incorporated into C++26 after extensive committee review, with the aim of simplifying SIMD programming.

However, recent benchmarks and analyses—such as those shared on Hacker News—show that std::simd compiles significantly slower, runs slower than scalar code, and defaults to suboptimal vector widths. Additionally, it cannot efficiently express certain operations critical for high-performance applications, such as fixed-width algorithms used in cryptography and codecs. In contrast, compiler auto-vectorizers in GCC, Clang, and MSVC have improved remarkably, often outperforming std::simd on key metrics.

Several open-source libraries have emerged that address the same problem space more effectively. Google’s Highway library, for example, offers runtime dispatch to select the best SIMD implementation dynamically, supporting scalable vectors like ARM’s SVE. Similarly, SIMDe provides portability for existing intrinsics, translating them across architectures without abstracting away fixed-width operations.

Why It Matters

This development matters because it highlights a disconnect between language standardization efforts and practical performance needs. While std::simd aims to simplify SIMD programming, its current implementation appears to lag behind the capabilities of modern compiler auto-vectorization and specialized libraries. For developers working in high-performance computing, multimedia, and cryptography, this raises doubts about adopting std::simd in production code.

Furthermore, the story underscores how the landscape of SIMD programming has evolved rapidly over the past decade, with hardware and compiler improvements surpassing what a library-based abstraction can provide. The inclusion of std::simd in C++26 may be seen as a missed opportunity to leverage these advancements fully.

Background

The concept of portable SIMD programming has been evolving since at least 2009, with Matthias Kretz’s Vc library pioneering the approach. The idea was to express data-parallelism through C++ types rather than intrinsics, leading to proposals like P0214 and P1928 that aimed to standardize such features. Over the years, compiler auto-vectorization has improved, and architectures like ARM’s SVE introduced scalable vectors, making fixed-width abstractions less relevant.

During this period, libraries such as Highway and SIMDe emerged, offering performance-portable solutions with runtime dispatch and intrinsics translation, respectively. These tools have been adopted in production environments, including Chromium, Firefox, and multimedia codecs, emphasizing their practical advantages over std::simd.

“Std::simd compiles 10x slower, runs slower than scalar loops, defaults to the wrong vector width, and can’t express the operations that actually matter in real SIMD code.”

— Hacker News user

“The original goal was to write SIMD code once and compile it everywhere. But the landscape has changed, and compiler auto-vectorization now often outperforms such abstractions.”

— Matthias Kretz

What Remains Unclear

It remains unclear how widely std::simd will be adopted given its current performance shortcomings. Developers and organizations may prefer existing libraries or rely on compiler auto-vectorization. The future updates or improvements to std::simd are also uncertain, as ongoing compiler optimizations could further diminish its relevance.

What’s Next

Next steps include monitoring updates from the C++ committee regarding std::simd, observing how compiler support evolves, and whether alternative libraries like Highway or SIMDe gain further traction. Developers may also experiment with these tools to determine the best approach for their specific use cases.

Key Questions

Why does std::simd perform worse than compiler auto-vectorization?

Benchmarks suggest that std::simd’s implementation introduces overhead and defaults to suboptimal vector widths, while modern compilers can often automatically generate more optimized code without the need for explicit abstractions.

Can std::simd be improved in future updates?

Potentially, but current community feedback indicates that fundamental architectural changes or compiler support improvements would be necessary to make std::simd competitive with existing solutions.

How does std::simd compare to libraries like Highway or SIMDe?

While std::simd offers a type-based abstraction, Highway provides runtime dispatch for performance portability, and SIMDe translates intrinsics for cross-architecture compatibility. Both are considered more practical in current high-performance contexts.

Is std::simd suitable for production use now?

Based on current benchmarks and community feedback, std::simd is not recommended for performance-critical production code. Developers should consider alternatives like Highway or rely on compiler auto-vectorization.