triton: MatMul tutorial fails for float32 inputs

I’ve ran the 3rd tutorial on matrix multiplication and the code runs successfully with the following versions

torch==2.0.0+cu118
triton==2.0.0

I then changed the dtypes of the input from half precision to single precision like below

a = torch.randn((512, 512), device='cuda', dtype=torch.float32)
b = torch.randn((512, 512), device='cuda', dtype=torch.float32)

Running this code then produces the output ❌ Triton and Torch differ

Why would making the inputs higher precision cause worse numerical error in the output? I’ve been observing this on other kernels which was how I ended up trying it on the tutorial as well.