NVIDIA.tex

\chapter{NVIDIA GPUs}
\label{chap:nvidia-gpus}


\section{C1060}
\label{sec:c1060}

Tesla C1060 is based on Tesla T10 GPU, which uses PCI Express (PCI-e) 2.0x16
bus with bandwidth: 

Maximum processing performance: 933 GFLOPs/sec (single precision).

RAM GDDR3 4GB with bandwidth: 102GB/sec; and I/O memory interface:
512-bit.

Number of cores: 240 (divided into 8 SM, each SM has 30 cores)


\section{Fermi (Tesla C20)}
\label{sec:fermi}

Fermi card is based on NVIDIA T20 GPU. NVIDIA plan to manufacture two
versions: C2050 and C2070. They use PCI Express (PCI-e) 2.0x16

Maximum processing performance: 520-630 GFLOPs/sec (double-precision).

RAM GDDR3 with ECC (error checking) support (can be enable/disabled by
end user). With ECC enabled, some memory is used for ECC bits. The I/O
memory interface is 384-bit. 

Number of cores: 448 (core clocks from 1.25GHz to 1.40 GHz, while with
Tesla is 1.296GHz)\footnote{previously, it was widely believed that it
  would feature 512 cores}. In turns, it consumes less energy (1.05V, compared
with 1.1875V in Tesla). 

References:
\begin{itemize}
\item \url{http://www.nvidia.com/docs/IO/43395/BD-04983-001_v01.pdf}
\end{itemize}
\subsection{C2050}
\label{sec:c2050}

RAM GDDR5 3G with bandwidth: 170 GB/sec. RAM decreases to 2.625GB, if
enable ECC.

\subsection{C2070}
\label{sec:c2070}

RAM GDDR5 6G with bandwidth: 6 GB/sec. RAM decreases to 5.25GB if
enable ECC.




%%% Local Variables: 
%%% mode: latex
%%% TeX-master: "gpucomputing"
%%% End: