inference/research
verifiedmedium · 0.80updated 2026-10-04

WebGPU (in-browser inference)

webgpu, wasm-simd

Fields

id
engines/local-webgpu
Repo
w3c/webgpu
Languages
WGSLJavaScriptC++Rust
License
W3C
First release
2023
Design goal
ubiquitous-hardwareminimal-depslocal/edge
Scheduling
none
Backends
webgpuwasm-simd
Formats
wgslonnxgguf
Notable features
WebGPU is a web standard for GPU compute and graphics, enabling in-browser LLM inferenceCompute shaders enable GEMM operations needed for transformer inferenceSupported in Chrome 113+, Safari 18+, Firefox 141+WGSL (WebGPU Shading Language) is the compute shader languageWebLLM and transformers.js are the primary LLM frameworks using WebGPUNo installation required - runs in the browser with GPU access
Hardware caveats
Performance is 2-5x slower than native GPU inference due to browser overheadMemory is limited by browser tab memory constraints (typically 2-4GB)No access to system RAM - all model weights must fit in GPU memory or be streamedBrowser support is recent - Safari 18+ (2024), Firefox 141+ (2025)No continuous batching or paged KV cache - single-stream inference onlyWGSL is less mature than CUDA/Metal - fewer optimized kernels
Best for
Zero-install LLM inference in the browserWeb applications that need local inference without server dependencyCross-platform deployment (Windows, macOS, Linux, ChromeOS) via browserSmall models (1-7B) that fit in browser memory constraints
Avoid for
Large models (70B+) - browser memory constraints are too tightMaximum performance - native GPU inference is 2-5x fasterProduction serving - no batching, scheduling, or multi-tenant supportNon-browser environments - WebGPU is browser-only
notes
WebGPU enables in-browser LLM inference via compute shaders. The primary frameworks are WebLLM (MLC) and transformers.js (Hugging Face). Performance is 2-5x slower than native GPU inference due to browser overhead and WGSL immaturity. Memory is limited by browser tab constraints (typically 2-4GB). WebGPU is the right choice for zero-install web applications and small models, but not for large-scale or high-performance local inference. ||| MAINTENANCE: NOT VERIFIED on 2026-10-04 - the engine record's `repo` field (w3c/webgpu) returned HTTP 404 from the GitHub API, so no maintenance fact could be pulled for it and none is estimated here. See [[ev-maintenance-status-is-four-dated-facts]].

Notes

WebGPU enables in-browser LLM inference via compute shaders. The primary frameworks are WebLLM (MLC) and transformers.js (Hugging Face). Performance is 2-5x slower than native GPU inference due to browser overhead and WGSL immaturity. Memory is limited by browser tab constraints (typically 2-4GB). WebGPU is the right choice for zero-install web applications and small models, but not for large-scale or high-performance local inference. ||| MAINTENANCE: NOT VERIFIED on 2026-10-04 - the engine record's `repo` field (w3c/webgpu) returned HTTP 404 from the GitHub API, so no maintenance fact could be pulled for it and none is estimated here. See [[ev-maintenance-status-is-four-dated-facts]].

Cross-references

This record references no other records — it is a leaf. It is cited by 2.