Having issues with MAX' Matmul on default Google Colab GPU (T4)

I am getting an error when trying to build a MAX Graph with a matmul on a T4 GPU (which is the default in Google Colab)

LLVM ERROR: Cannot select: intrinsic %llvm.nvvm.ldmatrix.sync.aligned.m8n8.x4.b16

Other Graph Ops work just fine. The full Notebook which creates the error: Google Colab (before execution: select runtime → T4)

Yes, we currently have Tensor Core support in the kernels for Ampere and newer NVIDIA GPUs, but Turing support was a new community addition and the Tensor Core operations haven’t been modified to extend support back to that architecture. Therefore, graphs that include matrix multiplication operations will fail on that architecture, but other GPU programming capabilities should still work.

Everything needed to extend support to Turing Tensor Cores should be present in the open source gpu module in the Mojo standard library, and we’d greatly appreciate any community contributions that could enable this.

1 Like

This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.