Wrapping your code with tf.device('/cpu:0'): I get 9.63 seconds (M1) and with GPU 60s.

Thanks! Filed feedback with the output and /tmp/...ld-snapshot contents. Bug number is 9839914. Yes, building for Apple silicon on Apple silicon. $ ghc --info ,("target platform string","aarch64-apple-darwin") ,("target os","OSDarwin") ,("target arch","ArchAArch64") ,("LLVM target","arm64-apple-darwin") ,("Build platform","aarch64-apple-darwin") ,("Host platform","aarch64-apple-darwin") ,("Target platform","aarch64-apple-darwin")

I tried -optl-mmacosx-version-min=11.0 and MACOSX_DEPLOYMENT_TARGET=11.0 which both resulted in many warnings like ld: warning: dylib (/opt/homebrew/lib/libpq.dylib) was built for newer macOS version (12.0) than being linked (11.0) followed by the same ld assertion failure.