\begin{table}%t5
\caption{\label{tab:code_efficiency}Computational efficiency of all codes/runs.}
%\centerline
 {\small \begin{tabular}{l c c c c c c c c}
\hline\hline
& \texttt{GADGET} & \texttt{PHANTOM~A} & \texttt{PHANTOM~B} & \texttt{VINE} & \texttt{ENZO} & \texttt{FLASH} & \texttt{TVD} & \texttt{ZEUS}  \\
\hline
architecture (CPU type)    & Itanium     & Clovertown  & Clovertown      & Itanium    & Itanium     & Xeon         & Xeon         & Xeon        \\
CPU clock rate             & $1.6$~GHz  & $2.66$~GHz & $2.66$~GHz     & $1.3$~GHz & $1.6$~GHz  & $1.6$~GHz   & $2.66$~GHz  & $2.33$~GHz \\
number of CPUs used        & 32          & 8           & 8               & 64         & 8 (32)      & 64           & 8            & 64          \\
parallelisation            & MPI         & OpenMP      & OpenMP          & OpenMP     & MPI         & MPI          & MPI          & MPI         \\
total CPU-h                & 6,490       & 248         & 248             & 15,000     & 165 (203)   & 256          & 10           & 315         \\
total CPU-h (norm.$^a$)    & 10,960      & 697         & 697             & 20,600     & 165 (203)   & 256          & 17           & 459         \\
\hline 
\end{tabular}}
\medskip
$^a$ CPU hours normalised to $256^3$ resolution elements, and normalised to a clock rate of $1.6~$GHz on an Intel Itanium/Xeon chip. These are very rough estimates that should only be accurate to within factors of a few, because of the different parallelisation hardware used on the various supercomputing platforms. Note however that \texttt{GADGET} and \texttt{ENZO} were run on the same supercomputing platform, and with the best run time and parameter optimisations that we could achieve on that supercomputer for both codes. The total CPU time for these two runs can thus be compared directly. \texttt{ENZO} was also run on 32 CPUs (values given in brackets).
\vspace*{3mm}
\end{table}