What it is

WebAssembly SIMD refers to Single Instruction, Multiple Data operations introduced to the WebAssembly specification. These operations define a portable subset of vector instructions that typically map directly to hardware-native instructions on the host CPU. The execution latency of those instructions varies across CPU models by physical circuit design, pipeline depth, and execution unit efficiency — a variation that can be measured from within a browser and used for hardware fingerprinting.

How the fingerprinting works

Browser engines (V8 in Chrome, JavaScriptCore in Safari) compile WASM SIMD operations to native machine code. For operations like complex byte permutations (i32x4_shuffle, i64x2_shuffle), CPUs without a dedicated hardware path emit a multi-instruction software emulation fallback. The latency gap between the hardware-native path and the emulation path is stable, hardware-specific, and measurable.

The measurement method: chain the SIMD operations in a dependency loop so each output feeds the next, then run that loop for millions of iterations. Dividing the total elapsed time by the iteration count reduces the impact of performance.now()’s 0.1ms quantization floor. The result is a per-operation latency profile that characterizes the underlying CPU.

Operations tested include floating-point (f32x4_div, f32x4_*), integer (i16x8_*), permutation (i8x16_shuffle, i64x2_shuffle), bitwise, and type conversion groups.

Empirical results (azerpas, May 2026)

Anthony Manikhouth built and evaluated this system across 2,124 unique devices: 1,501 mobile, 623 desktop. Source platforms: BrowserStack real devices, AWS/GCP/Vast.ai bare metal.

Classification model: K-Nearest Neighbors (k=5) with distance weighting, using per-operation timing vectors as features. Evaluation method: Leave-One-Out (LOO) cross-validation.

Results after iterative model improvements (filtering CPUs with <10 training runs, using fastest-raw-sample per op instead of median, grouping Apple Pro/non-Pro variants):

EngineCPU accuracyBrand accuracy
V8 (Chrome)82.1%95.2%
JSC (Safari)~80%100%

Per-CPU highlights for V8:

  • AMD Ryzen 5 9600X: 100% (29/29)
  • Snapdragon 8 Elite: 94% (247/264)
  • Google Tensor G2: 58% (73/125) — shares architecture with G3, frequently confused
  • Intel Core i7-7700K: 25% — low training count, not representative

PCA of V8 runs shows Intel and AMD clusters clearly isolated, with the first two principal components capturing 92% of variance. t-SNE distinguishes AMD EPYC generations from 2021 to 2025. Apple A19 family is well isolated from earlier chips in JSC runs; A16 and M2 (both 2022) cluster closely due to architectural overlap.

Spoofing and stability

Performance.now() jitter: injecting noise into the timer would appear as an unknown trace rather than a known CPU profile, which is counterproductive for an attacker trying to impersonate a real user’s device.

Relative ratio robustness: the signal relies on ratios between dozens of operation latencies, not absolute timing. Random jitter on start/end timestamps does not destroy the ratio structure.

Compiler drift: if V8 optimizes its i8x16_shuffle lowering, the baseline shifts overnight. Stable fingerprinting requires operations that map 1:1 to native instructions across all engines, minimizing the compilation layer between measurement and hardware.

Dynamic challenge defense: a server could issue a parameterized benchmark challenge that forces the client to recompute the trace for a specific operation set, preventing pre-harvested replay.

Engine variance

The same physical hardware yields different timing profiles depending on the engine. On an M3 MacBook, f32x4_div varies by ~25% between V8 and JSC. Any classifier must be engine-specific; a model trained on V8 data does not generalize to JSC runs.

Implications for scraping

A fingerprinting system using WASM SIMD timing can identify the CPU model powering a scraper with ~82% accuracy. Device farms (AWS, GCP, Vast.ai) expose their server CPU models (AMD EPYC, Intel Xeon generations) through this signal. A scraper claiming to be a consumer iPhone while running on a Xeon would be contradicted by the SIMD timing profile.

This technique is not yet deployed in production bot detection stacks as of May 2026 — it is at the research/PoC stage. The author is collecting organic telemetry from real users to refine the training data.

TWSC experience

Not tested by TWSC. Research documented here is from Anthony Manikhouth (azerpas.com, May 2026).

Sources


The Web Scraping Wiki is compiled from the tests and research published on The Web Scraping Club, the newsletter for web scraping professionals. Subscribe to get the next article in your inbox.