Help with optimization/profiling

You can speed it up a lot by loading the data more directly without using binary:

{-# INLINE unconsW8 #-}
unconsW8 :: BSL.ByteString -> Maybe (Word8, BSL.ByteString)
unconsW8 = BSL.uncons

{-# INLINE unconsW16 #-}
unconsW16 :: BSL.ByteString -> Maybe (Word16, BSL.ByteString)
unconsW16 bs = do
  (x, bs') <- unconsW8 bs
  (y, bs'') <- unconsW8 bs'
  pure (fromIntegral x .|. (fromIntegral y `shiftL` 8), bs'')

{-# INLINE unconsW32 #-}
unconsW32 :: BSL.ByteString -> Maybe (Word32, BSL.ByteString)
unconsW32 bs = do
  (x, bs') <- unconsW16 bs
  (y, bs'') <- unconsW16 bs'
  pure (fromIntegral x .|. (fromIntegral y `shiftL` 16), bs'')

{-# INLINE unconsFloat #-}
unconsFloat :: BSL.ByteString -> Maybe (Float, BSL.ByteString)
unconsFloat bs = case unconsW32 bs of
    Just (x, xs) -> Just (castWord32ToFloat x, xs)
    Nothing -> Nothing

main = do
    pixelsRaw <- BSL.readFile "matrix.bin"
    let pixels = SV.unfoldrExactN (5000*5000) (\s -> case unconsFloat s of Nothing -> (0,s); Just (x, s') -> (x, s')) pixelsRaw
    BSL.writeFile "/tmp/twotime-bsl.svg" (pageSvg pixels)

That also removes all GC overhead:

before: (click for more details) after: (click for more details)

Edit: SV.unfoldrExactN is much better than SV.replicateM