English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

WebAssembly 3.0 In-Depth Report: Technically Solid, Real-World Adoption Far Smaller Than Claimed

Forum topic · ✨步子哥 · 2026-08-06

Summary

This in-depth technical audit of WebAssembly 3.0 concludes that while the spec is technically honest and well engineered, its real-world footprint is roughly two orders of magnitude smaller than commonly believed. Key findings: real WebAssembly usage accounts for only 0.0613% of Chrome page loads, contradicting the widely circulated "5.5%" claim (a 90x error). Chrome telemetry shows the flagship Memory64 feature at 0.00000146% of page loads and declining even after it shipped by default; the new exnref exception handling is used about 1/28,900 as often as the legacy scheme. The report documents that no standalone "wasm-core-3" document exists on the W3C site (3.0 lives inside a document short-named wasm-core-2 under an evergreen model), that two official documents give conflicting feature lists for 3.0, and that the Profiles feature entered the 3.0 text despite being only a Phase 1 proposal. Verified data includes engine support matrices (Chrome 137, Firefox 134, Safari 26, Node.js 25 for full 3.0), V8 source-code analysis showing Memory64's actual limits are 16 GiB on 64-bit hosts and identical to memory32 (2 GiB) on 32-bit hosts, and multi-memory implementation caps differing 1000x between engines (V8: 100,000; SpiderMonkey/JSC: 100). The report also debunks numerous viral claims, including "16 EB of usable memory", "Kotlin/Wasm 50-70% smaller binaries", and a nonexistent "Safari 145".

WebAssembly 3.0: A Full-Perspective Deep-Dive Report

One-line conclusion: WebAssembly 3.0 is a technically honest, well-engineered standard whose narrative does not survive scrutiny; its real-world presence is about two orders of magnitude smaller than almost everyone assumes.

The type-system overhaul (typed function references + WasmGC) is the most important foundational work since Wasm's MVP and is technically beyond reproach. But the numbers tell a different story: no document named wasm-core-3 exists on the W3C site; two official documents give contradictory answers on what 3.0 contains; the flagship Memory64 feature accounts for 0.00000146% of all Chrome page loads and declined after being enabled by default; the new exnref exception handling is used about 1/28,900 as often as the legacy scheme (99.9965% of real-world exception handling still runs on legacy). And the widely circulated "5.5% of Chrome pages use WebAssembly" claim has a true value of 0.0613% — a 90x gap.

Key points

Version identity: what "Wasm 3.0" actually is

  • 3.0 was declared complete on 2025-09-17 as a new "live standard"; claims of other dates are wrong.
  • The 3.0 core text physically lives inside a Candidate Standard document whose short name is still wasm-core-2. There is no standalone wasm-core-3 document, no errata, and no citable frozen version — under the evergreen model adopted after 2.0 (per spec editor Andreas Rossberg, it will "without ever technically move to the final Recommendation state").
  • The proposal tracking table lists 11 features for 3.0; the core spec change history lists 10, with three contradictions: Branch Hinting and JS String Builtins are marked core in the table but absent from the core change history, while Profiles (still Phase 1, championed by the spec's own editor) appears in the 3.0 text and is the only entry with no proposal-directory footnote.
  • Adoption: three incompatible metrics

  • Chrome page loads: 0.0613% actually instantiate Wasm (UseCounter, 2026-08-04). Sites: 0.35% desktop / 0.28% mobile request a .wasm resource (Web Almanac 2025). All-site samples: 0.05–0.1%. These measure different things and cannot be interconverted.
  • Debunked numbers: "5.5%" (90x error), "4.5%", "41%", "16 EB usable memory" (the spec ceiling is 16 EiB verification-level; users reach at most 16 GiB, less in practice), "Kotlin/Wasm binaries 50–70% smaller" (no official source), "Coremark 95% of native", "Fermyon 75M req/s", and the nonexistent "Safari 145".
  • Overall usage grew only ~1.20x year-over-year from a base of 0.05%; no inflection point appears around the 3.0 announcement. At the current slope, Wasm would need ~15 years to reach 1% of page loads.
  • Feature reality check (Chrome telemetry + engine source)

  • Memory64: V8's wasm-limits.h states page counts "are still a 32-bit number for now" (128 TB ceiling); on 32-bit hosts memory64 and memory32 have identical limits (32,767 pages = 2 GiB). Its co-champion: "The only reason to use Memory64 is if you actually need more than 4GB... We still don't even have a way to release memory back to the operating system." Usage today is 18% of its flag-era peak. The companion table64 is still WIP in V8 and Firefox.
  • Exception handling: the design looped back to exnref six years after it was first introduced (spec issue #281), with the authors themselves writing "deprecation may not be feasible". Median telemetry: legacy EH 0.00607% vs exnref 0.00000021% — ~28,900x gap; 99.9965% of real-world Wasm exception handling runs legacy. Kotlin/Wasm officially still requires "legacy exception handling".
  • Relaxed SIMD and tail calls" "growth": apparent hundreds-fold year-over-year jumps are single-day step changes from a single deployer, not adoption curves.
  • WasmGC: correctly understood, it lets modules declare struct/array types on the host GC heap — not "a GC added to Wasm". Its usage curve is flat year-over-year (1.09–1.2x), and three independent counters (TypedFuncRef, JS String Builtins, WasmGC) share identical peaks and values to the 5th decimal — strong inference that usage is essentially one large deployer (likely J2CL-compiled Google properties). Meanwhile the non-standardized Custom Descriptors went from zero to 0.0051% in eight months — 4.5x faster than WasmGC — proving low 3.0 usage reflects lack of necessity, not lack of time.
  • Multi-memory: engine caps differ 1000x (V8: 100,000; SpiderMonkey and JSC: 100). Cross-engine safe value is 100; compliance test suites will not catch violations.
  • Deterministic Profile × Relaxed SIMD: the DET profile pins every relaxed instruction to a fixed spec-defined behavior — an elegant institutional solution for reproducibility. In practice, only Wasmtime marks support, with zero published case studies.

What is NOT in 3.0

Threads (Phase 4), Shared-Everything Threads (Phase 1), Stack Switching (Phase 3), JSPI (standardized but explicitly not 3.0), Component Model, Custom Descriptors (Stage 3) —
3.0 adds nothing in the concurrency/control-flow domain**. Web Wasm threads remain largely locked out by cross-origin isolation requirements after Spectre/Meltdown.

Engine support for full 3.0

Chrome 137, Firefox 134, Safari 26 (year-based versioning), Node.js 25 — 18 major Chrome versions above WasmGC's first availability in Chrome 119.

Stated limitations of the report

Chrome telemetry covers only Chrome/UMA users; server-side Wasm (Wasmtime, Fermyon, Fastly, Cloudflare Workers) is entirely outside these statistics — if Wasm's main battleground has shifted server-side, this data systematically understates its importance. UseCounter counts "ever triggered per page load", not intensity, and provides no attribution.

Conclusion

Wasm 3.0's technical content is solid, particularly the typed-references and GC foundation; the criticism lands on narrative and adoption. The version number functions more as a marketing snapshot than a rigorously defined release, and real-world traffic data shows 3.0's features have barely registered — while the features big vendors actually ship in production (standardized or not) are what move the curve.

*Methodology note: low-frequency counters (<0.001%) use 30-day medians; the last 3 days of series are treated as unpublished; all bucket IDs are cross-verified against Chrome's official feature-popularity dataset and V8 source.*

Tags

#webassembly#wasm-3.0#chrome-telemetry#wasmgc#memory64#exception-handling#browser-engines#specification-analysis

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178603046