N
vidia has launched an effort
to develop an entirely new
customer base for its graphics
processors. Dubbed Tesla, the new
product line consists of technology
to build supercomputers designed
for scientific and industrial applica-
tions using existing GPUs.
Tesla will leverage the com-
pany’s existing top-end graph-
ics processor, currently the G
chip that forms the basis of the
consumer-oriented GeForce 8800
GTX, to produce massively paral-
lel workstations, desktop super-
computers, and rack-mount GPU
servers. The G80s used for Tesla
and Quadra products differ from
those used to produce 8800 cards
in one key way: They’re capable of per-
forming double-precision (64-bit) floating-
point operations. This feature is removed
from the dies destined for consumer prod-
ucts. Although Tesla won’t have consumer
applications, it will be interesting to watch
a cluster of relatively inexpensive GPUs
take on the likes of Cray and IBM.
Application developers will use
Nvidia’s CUDA (Compute Unified Device
Architecture) C-language toolkit to create
custom software programs to run on Tesla
products. Nvidia first announced CUDA in
November 2006, which gave developersa head start using Nvidia’s Quadra line of
GPUs. Evolved Machines, based in Palo
Alto, CA, for example, is developing a
Tesla-based system for reverse-engineer-
ing neurons in the human brain. Its prod-
uct is being used to engineer devices that
can not only sense and recognize objects
and odors but also learn about new
objects and odors. The company had been
using x86 microprocessors to run its simu-
lations but found that Nvidia’s technology
could run them 130 times faster.
As the general manager of Nvidia’s
computing group, Andy Keane, explains it,such dramatic increases in performance
are possible because “a CPU has an
incredible amount of technology designed
to perform a certain type of work, but it’s
designed to handle one task at a time and
to switch from one task to another very
quickly. A CPU has large caches and other
infrastructure designed to keep its execu-
tion pipeline full. GPU computing is very
different. It’s designed to perform tasks in
parallel with a large cache of memory. We
view GPU computing as being comple-
mentary to CPU computing.”
Tesla is debuting in three forms: The
Tesla C870 ($1,500) consists of a single
8-series GPU on a PCI Express card. The
Tesla D870 ($7,500) features two 8-series
GPUs in an external chassis. The top-
of-the-line Tesla S870 ($12,000) houses
four passively cooled 8-series GPUs in a
rack-mount chassis that will also house
a microprocessor to enable external
management and 40 56mm cooling fans.
Nvidia eventually plans to offer dual-GPU
cards to enable 8-GPU server products.quickstart The beginning of The magazine, where arTicles are small
Nvidia Introduces a
Supercomputer Line
08 MAXIMUMPC september 2007
New GPU products
aimed at science
and industry
Boston Scientific is using a Tesla-based supercomputer to design biomedical implants that
are safe for use with diagnostic imaging tools such as MRIs.SUPERCOMPUTER ARCHITECTURE COMPARISON
SUPERCOMPUTER MANUFACTURER PRIMARY FLOATING-POINT-
PROCESSORS OPERATIONS/SEC
Cray XT4 Cray AMD Opteron 5.6 to 318 TFLOP
IBM Blue Gene/P IBM IBM PowerPC 450 13.9 TFLOP to 1.0 PFLOP
Nvidia Tesla Nvidia Nvidia 8-series 500 GFLOP to 2.0 TFLOP