Tensorflow on open source GPUs
summary
TensorFlow builds to CUDA/PTX currently, so work focuses on replacing CUDA with a SYCL/DPC++ frontend that emits SPIR-V kernel vs shader binaries, feeding Mesa’s NIR pipeline and driver backends. Eigen and a SYCL backend are merged but TensorFlow GPU support is still in progress; Clover OpenCL work stalls on unstructured control flow so LLVMpipe gets extended to run SPIR-V testing. Intel’s Level Zero runtime and a DPC++/SYCL → SPIR‑V toolchain exist; a Vulkan/LLVMpipe software renderer approaches conformance but games aren’t running yet.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:13Outline
- 1:25Tensorflow
- 3:15ICON
- 4:14Building
- 7:24Backend
- 9:24Execution
- 12:20OpenCL
- 14:50Vulcan
- 17:28Shader vs Kernel
- 21:27One API initiative
- 23:33Sickle
- 24:42Level Zero
- 26:56Current State
- 31:09Malcolm in the Middle
- 32:40Clover OpenCL
- 40:09Conclusion
related talks
28mTensorFlow 1.0: open source machine learning
25mThe status of Tensorflow on Arm-based HPC
26mMachine Learning with TensorFlow
25m"Rust for AI & Accelerated Computing"
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
40mIntroduction to Swift for Tensorflow
20mTensorFlow.js Bringing Machine Learning to the Web and Beyond
39mSwift for TensorFlow: Machine Learning with No Boundaries