I am often asked which NVIDIA products to buy for
scientific computing using CUDA. My current advice is:
a) For initial development, start with a PC / workstation with a 1GB or
2GB GTX560 card; this uses the latest Fermi GPU with L1/L2
caches, uses a maximum of 170W, will fit into most PCs (though it
is a long PCIe card and two slots wide), and costs around £200
from suppliers such as
dabs.com.
For the best performance, put it in a workstation with a 16x PCIe
gen 2 slot.
Note that the GTX560 is quite big (which could be a problem in a
small computer case) and requires two 6-pin connectors (see notes
below about this). If this is a problem, then one can start with
the GT440 card (£70 from dabs.com)
which is not as long, and appears not to use any 6-pin power connectors.
b) If double precision floating point arithmetic is important, consider the new
Quadro 4000
which has much better double precision performance than the GTX cards in which
the double precision performance is only 1/8 of the single precision performance.
c) For production computing, move to
S2050
Tesla servers; these have 4 GPUs in a 1U server, each with 3GB graphics
memory. They also use the Fermi GPU and are produced by NVIDIA with tighter
quality control, but note that they do not provide any video output.
I suggest using the SuperMicro
Twin2 as the host system; each of its four independent compute nodes
can connect to half of a S2050, and they also come with QDR Infiniband
to form a tightly-coupled GPU compute cluster.
See also Supermicro's new
GPU servers in which two Tesla GPUs are housed within a 1U server.
d) HP has the
HP SL6500 Scalable System
which is a 4U unit which can hold 4 compute nodes, each with 3 M2050/2070 GPUs and
QDR Infiniband.
e) Dell has the
PowerEdge C410x PCie expansion chassis
which is a 3U unit which can hold up to 16 GPUs to be connected to 1-4 separate
compute nodes.
Power connectors for NVIDIA cards
All of NVIDIA's high-end cards require one or more PCI Express power
connectors. These come in two flavours, 6-pin and 8-pin. Some low-end cards
require just one 6-pin connection, while the most powerful card requires
two 8-pin connections. It used to be difficult to find out what each card used,
but now you can find out by going to this
GeForce family
page, selecting the card of interest, and then the Specifications tab, and
then scrolling to the bottom.
| Card |
PCIe power connectors |
| GTX 550 |
6-pin |
| GTX 560 |
2 x 6-pin |
| GTX 590 |
2 x 8-pin |
| Tesla C2050 |
8-pin or 2 x 6-pin |
Most new power supplies offer at least one 6-pin connection, and quite a
few provide an 8-pin connection.
A lot of graphics cards come with adapters to convert the connections
offered by an existing power supply to what you need. In particular,
the GTX560 may come with an adapter like
this
to convert two 4-pin molex supplies (for disk drives) into one 6-pin
PCI Express connector. This enables you to install the GTX560 in a
system which has a power supply with only one 6-pin connector.