{"id":10062,"date":"2023-10-12T16:18:48","date_gmt":"2023-10-12T07:18:48","guid":{"rendered":"https:\/\/www.stechstar.com\/user\/wordpress\/?p=10062"},"modified":"2023-10-12T16:18:48","modified_gmt":"2023-10-12T07:18:48","slug":"%ec%9d%b8%ea%b3%b5%ec%a7%80%eb%8a%a5-%ea%b8%b0%ec%88%a0-webgpu-powered-machine-learning-in-the-browser-with-apache-tvm-apache-tvm%ec%9d%84-%ec%82%ac%ec%9a%a9%ed%95%98%ec%97%ac-%eb%b8%8c%eb%9d%bc","status":"publish","type":"post","link":"https:\/\/www.stechstar.com\/user\/wordpress\/%ec%9d%b8%ea%b3%b5%ec%a7%80%eb%8a%a5-%ea%b8%b0%ec%88%a0-webgpu-powered-machine-learning-in-the-browser-with-apache-tvm-apache-tvm%ec%9d%84-%ec%82%ac%ec%9a%a9%ed%95%98%ec%97%ac-%eb%b8%8c%eb%9d%bc\/","title":{"rendered":"[\uc778\uacf5\uc9c0\ub2a5 \uae30\uc220] WebGPU powered machine learning in the browser with Apache TVM : Apache TVM\uc744 \uc0ac\uc6a9\ud558\uc5ec \ube0c\ub77c\uc6b0\uc800\uc5d0\uc11c WebGPU \uae30\ubc18 \uae30\uacc4 \ud559\uc2b5"},"content":{"rendered":"

[\uc778\uacf5\uc9c0\ub2a5 \uae30\uc220] WebGPU powered machine learning in the browser with Apache TVM : Apache TVM\uc744 \uc0ac\uc6a9\ud558\uc5ec \ube0c\ub77c\uc6b0\uc800\uc5d0\uc11c WebGPU \uae30\ubc18 \uae30\uacc4 \ud559\uc2b5<\/h1>\n

By Tianqi Chen, Jared Roesch, and Jason Knight<\/mark><\/span><\/p>\n

TL;DR<\/b><\/mark><\/span><\/p>\n

Easy ML deployment in the browser with near-native GPU performance and WebAssembly<\/mark><\/span><\/p>\n

We introduced support for WASM and WebGPU backends to the <\/mark><\/span>Apache TVM deep learning compiler<\/mark><\/span> Our initial experiments shows that TVM\u2019s WebGPU backend can get <\/mark><\/span>close to native GPU performance<\/b><\/mark><\/span> when deploying models in the browser.<\/mark><\/span><\/p>\n

Introduction \u2014 faster browser applications<\/b><\/mark><\/span><\/p>\n

WebGPU<\/mark><\/span> and <\/mark><\/span>WebAssembly <\/mark><\/span>are two up-and-coming browser standards aimed at making browser based applications as efficient as native applications while retaining hardware portability. WebGPU first builds on top of the lower level graphics API movements to enable more full featured GPU accelerated compute capabilities as compared to its predecessor WebGL, and WebAssembly (WASM) is a similar effort for unlocking CPU based compute capabilities. And best of all, WASM is most likely <\/mark><\/span>already enabled<\/mark><\/span> in the browser you\u2019re reading this in, with WebGPU <\/mark><\/span>coming soon<\/mark><\/span>.<\/mark><\/span><\/p>\n

For some of the exciting capabilities and use cases behind WASM particularly, please check out <\/mark><\/span>Lin Clark\u2019s excellent WASM \u201cskill tree\u201d<\/mark><\/span><\/p>\n

<\/span><\/p>\n

WASM\u2019s post MVP skill tree<\/mark><\/span><\/p>\n

Leveraging more of these WASM capabilities such as SIMD support and multi-threading as they mature is an exciting area of future work, but let\u2019s see how far we can get with WASM and WASI (standard interfaces) to feed WebGPU compute shaders.<\/mark><\/span><\/p>\n

Machine learning in the browser<\/b><\/mark><\/span><\/p>\n

While efforts like TensorFlow.js and ONNX.js attempt to bring machine learning to the browser, there still exist non-trivial gaps in performance between the web versions and native ones. One of the many reasons behind these gaps are the lack of standard and performant access to the GPU on the web. WebGL lacks important features such as compute shaders and generic storage buffers that are necessary for high performance deep learning.<\/mark><\/span><\/p>\n

WebGPU offers the potential to bridge this gap with its first-class compute shader support.<\/mark><\/span><\/p>\n

To explore the potential of using WebGPU for deploying machine learning in the browser, we enhanced the deep learning compiler <\/mark><\/span>Apache (incubating) TVM <\/mark><\/span>to target WASM (for host code that computes the launching parameters and calls into the device launch) and WebGPU (for device execution). Our preliminary results are quite positive \u2014 for the first time, we can deploy machine learning applications on the web while still getting near native performance on the GPU.<\/mark><\/span><\/p>\n

Please check out our <\/mark><\/span>companion Apache TVM technical blog post<\/mark><\/span> for the technical details, but in short, we were able to achieve near native GPU performance in the browser through WASM and WebGPU by combining:<\/mark><\/span><\/p>\n