Nvidia announced today that Dell has begun shipping workstations with their Nvidia Tesla C1060 supercomputer card. Dell Precision R5400, T7500 and T5500 all offer Tesla GPU compute cards, which, according to Greg Weir, Senior Manager at Dell, “put the power of supercomputing on the desktop”.
Dell’s R5400 is a rack-mountable workstation computer designed for cluster operations. The T5500 and T7500 are tower-model workstations which sell for around $1800 base, plus the add-ons, which now include the C1060 which has an MSRP base price of around $1599.
The Tesla C1060 GPU provides raw computing power which is roughly equal to a small- to mid-range traditional CPU supercomputer cluster, though in a traditional dual-slot graphics card-sized configuration. The workloads it can operate on are not always the same as full-blown supercomputers, which can handle anything. Therefore, the C1060 does have real limitations (as do all graphics-based compute engines).
While the Tesla C1060 looks like a graphics card in appearance card, and even appears to software that way, it has no graphics port outputs. The Tesla is designed exclusively to be a compute engine which operates inside your PC.
Nvidia introduced double-precision (64-bit) floating point compute abilities with Tesla. The card itself is similar to high-end graphics cards offered by Nvidia, capable of nearly a Teraflop of dedicated single-precision compute performance (933 GFlops), with something around 80 Gigaflops in double-precision, though it is fully IEEE-754 compliant. It consumes up to 188 watts, has a memory bandwidth of 102 GB/s on a 512-bit interface 800 MHz clock, with a 1.3 GHz core clock with 240 stream processors. Each card comes with 4 GB of GDDR3 memory. Even though the Tesla C1060 card itself consumes up to 188 watts each, the reality is its compute abilities far exceed a traditional CPU-based supercomputer cluster, and for a fraction of the power and, according to Nvidia, for 1/100th the price.
The card’s dedicated compute is typically targeted at the oil and gas industries, computational finance, fluid dynamics, medical research and weather modeling — jobs which are traditionally sifting through volumes of data, computing with common operations like add, sub, multiply and divide.
It operates with Nvidia’s CUDA programming language, which exposes the compute abilities of the card through a C-like software interface.
See Nvidia’s press release, and information page about the Tesla C1060 released earlier this year in January. Nvidia also has a Tesla Personal Supercomputer Store, which shows several dedicated clusters which contain a series of Tesla cards inside, and are sold as true personal supercomputers. TigerDirect also published a YouTube video on Nvidia’s Tesla cards.
Rick’s Opinion
Nvidia’s second-generation Tesla C1060 was introduced nearly a year ago. That card is not up to today’s levels of performance and even greater performance can be obtained today through Nvidia or ATI’s highest-end graphics cards.
In fact, the Tesla is an expensive parallel compute engine, and one not required for massively parallel compute abilities thanks to the advancements in CUDA and BrookGPU libraries for Nvidia and ATI graphics cards.
Buying two or more high-end Nvidia’s GTX 2xx or ATI Radeon HD 4xxx series cards might be a wiser decision, as these cards yield well over 1 Teraflop of performance today on less power, while also providing usable graphics abilities for the workstation. And with Nvidia’s SLI and ATI’s CrossFire, the cards can be joined together for a single-instance compute appearance with even greater resources and power. (story Link)