3 ms·

Thanks for the quick reply! About hardware support, I was wondering if the LPU has a hardware instruction to compute the attention matrix similar to the MatrixM

Thanks for the quick reply! About hardware support, I was wondering if the LPU has a hardware instruction to compute the attention matrix similar to the MatrixMultiply/Convolve instruction in the TPU ISA. (Maybe a hardware instruction which fuses a softmax on the matmul epilogue?)

tome 3y ago
We don't have a hardware instruction but we do have some patented technology around using a matrix engine to efficiently calculate other linear algebra operations such as convolution.