Santa Clara (CA) – Nvidia today announced Tesla, a third product line next to the GeForce and Quadro graphics products. The company aims to use Tesla cards and the massive floating point horsepower of its graphics processors to take over a portion of the lucrative supercomputing market.The core of each Tesla device is a GeForce 8-series GPU as well as the general component layout of the high-end Quadro FX 5600 workstation graphics card with 1.5 GB of memory (in Tesla, it has 1.35 GB). The only noteworthy difference between the FX 5600 and a Tesla card is the fact that the supercomputing-targeted devices lack the graphics outputs on the backpanel, which we were told, allows Nvidia to increase the clock speed on Tesla.
While the actual clock speed of the Tesla GeForce GPU is kept under wraps, Nvidia said that one processor (used in the C870 add-in card) is good for a performance of 518 GFlops, two processors (used in the deskside supercomputer D870, which integrates two C870 cards) will bring 1 TFlops; the Tesla GPU server with four processors will hit 2 TFlops.
In terms of pure number crunching horsepower, Nvidia told us that one GeForce GPU can match the combined performance of 40 x86 processors. In addition to the raw performance, Tesla also makes a case for power efficiency: The C870 is rated at a maximum power consumption of 170 watts and the GPU server at 800 watts, which may sound a lot at first look. However, 40 low-power x86 processors would run at a typical 1600 watts. With a common power budget of about 25 kilowatts per rackserver, a Tesla GPU server rack has a theoretical maximum performance of more than 60 TFlops – which would put the floating point rating of such a device among the 15 fastest supercomputers currently ranked on the Top 500 Supercomputer list.
Similarities to ATI’s stream processor card, implications for developers
Readers, who have been following recent general purpose GPU announcements, will remember that ATI has product in its portfolio that is very similar to the Tesla C870 – the stream processor card (which is based on a R580 GPU and 1 GB of memory). Both products follow the same concept to make the massively processing capability provided by shader processors available to run arbitrary code instead of graphics code.
Developers such as John Stone and James Philips, senior research programmers at the Beckman Institute of Advanced Science and Technology at the University of Illinois, have been looking at accelerators such as GPUs for some, but have been limited mainly by bugs in shader drivers. Stone told us that much of his work with GPUs in the past was focused “on finding driver bugs” and “writing his applications around them” in order to make the technology usable for scientific simulations. “There can be a lot of rounding errors and because of this very fact, I wasn’t very excited about working with GPUs,” he said.
However, both AMD and Nvidia came up with a programming model to solve this problem. On AMD’s side, it is called CTM (“close to metal”) and on Nvidia’s side it is CUDA (“Compute Unified Device Architecture”). At this time, it appears to come down to personal liking which model is preferred by a developer, as, for example, there are some universities that are working with CTM (such as Stanford’s Folding@Home project) and there are some that are working with CUDA. Stone and Philips are focusing on the Nvidia model as they claim its C++-based language model is easier to deal with than AMD’s CTM version, which uses a low-level assembly language.
While CUDA works very much like a regular programming model and, according to Stone, can deliver results very quickly, the big challenge in exploiting these devices will be knowledge to write advanced parallelized code for these GPGPUs. Stone believes that especially coders who have written code for (massively parallel) supercomputers before will have an easy transition opportunity. Of course, knowledge of the hardware, graphics processing and a good look at the parallelizable parts of applications help to take advantage of the technology.
Shane Ryoo, a graduate research assistant at the University of Illinois at Urbana-Champaign, said that CUDA will allow programmers with some experience in developing threaded applications to get “really good results right off the bat.” However, it will be the fine-tuning process, which will increase the value of GPGPUs: Ryoo noted that expert knowledge that will allow developers to squeeze the best possible performance out of GPUs, sometimes can accelerate application code by a factor of 5x or greater.
Nvidia is well aware of this challenge and has begun assisting universities in establishing classes and developing course material focusing on massively parallel programming and CUDA in particular. Eventually, the company hopes, that GPGPU programming will become a standard part in computer science course work and help to educate a whole new generation of programmers. So far, Nvidia has taught courses at the University of Illinois, The University of California, the University of North Carolina and Purdue University. Nvidia said that several universities are developing their own courses, including the University of Virginia, the University of Pennsylvania, Oregon State University, the University of Wisconsin. Caltech, MIT, Berkeley and Stanford have been offering “legacy” GPGPU and GPU programming classes, according to Nvidia chief scientist David Kirk.
Read on the next page: The payoff, cost and the impact on the consumer
The payoff: Accelerated applications
If the capabilities of these GPGPUs are exploited, there can be a big payoff. Stone, who is working on Nanoscale Molecular Dynamics (NAMD) as well as Visual Molecular Dynamics (VMD), said that a virus simulation that took 110 CPU hours on a SGI Altix Itanium 2 supercomputer at NCSA required only 27 GPU minutes on a GeForce 8 graphics processor – which translates into a 240x speedup.In an example that showcases an impact that can touch many lifes, Ryoo and his team are working on an interactive, medical MRI application that substantially increases the resolution of MRI scans thanks to the added processing power. As a result, they expect to be able to deliver much finer images, which allow physicians to detect tumors at an earlier state or differentiate between a blip or an actual tumor.
In a demonstration showed during an Nvidia event, a representative from Headwave, a company that provides geophysical data analysis, highlighted a 4D application, which allows users to visualize gigabytes and apparently even terabytes of data in a three-dimensional scale and even apply a time filter to display changes to geological layers over time. The company claims that GPUs are accelerating their application by about 2000% and are delivering an output of about 2000 MB/s.
In fairness, we should mention that Tesla (or stream processor cards for that matter) will not be able to replace supercomputers, which continue to provide a memory bandwidth a few Tesla cards cannot match. Scientists such as Stone believe that products such as Tesla will make their way into supercomputers to create an overall more balanced environment. “Number crunching was the limiting factor up until now. Now Infiniband will be a problem,” he said.
GPGPUs are likely to have a greater impact on deskside supercomputers in the short term. While scientists today have to apply for expensive supercomputer time and in most cases have to wait several days until their application can be processed - if those requests are not turned down anyway – there is now an opportunity to run many of those tests on a desk right in the lab. Conceivably, GPGPUs will allow more scientists to run more and higher quality simulations in less time.
Cost and impact on the consumer
Nvidia’s Tesla products will start at $1500 for the single GPU add-in card; the 2-GPU deskside unit will run for $7500 and the 4 GPU server, which soon will also be offered in an 8 GPU version, will sell for $12,000. Leaving out of consideration that, at least to our knowledge, Tesla is not yet available, these apparently lofty price tags turn out to be bargains at a closer look.
The C870 not only undercuts the ATI stream processor card, which currently sells for about $2000, but also Nvidia’s own workstation products. The C870, at $1300, compares to a Quadro FX 5600 graphics card, which requires and investment in the neighborhood of $3000 and up. Clearspeed’s CSX600 accelerator card, which provides a performance of about 100 GFlops, is selling in volume for about $7500.
A representative of Evolved Machines told us that the company plans to be offering a 12 TFlops Tesla server, which will cost somewhere between $60,000 and $70,000, but will be fast enough to match the floating point performance of the 19th fastest supercomputer on the Top-500 list.
Stone told us that even if the GPUs per se may appear to be expensive for a consumer point of view, they “are available for far less money than the next best thing that is available today.”
So, what does that mean for the consumer? Clearly, there is only an indirect benefit for most consumers that we may see in improved research results down the road. However, as all technologies, these GPUs will get cheaper over time and even today, a $1300 card would be in reach for enthusiasts, who often spend substantially more than $5000 on their rig. The fact is that there is no magic necessary to make these cards work on a PC - and CUDA even works with GeForce 8 graphics cards, which can be had for less than $250 in the case of 8600-series models. The real question is: When will there be applications that take advantage of this technology and will they provide enough incentive for consumers to purchase a GeForce 8 card? Industry experts believe that it will be up do developers to come up with new applications that will take advantage of the capability of GPGPUs on the desktop.
Nvidia CEO Jen-Hsun Huang told TG Daily that Tesla will be strictly focused for the enterprise market and will not be making its way to the consumer market. In the end, it will be up to the GeForce product groups to leverage CUDA on desktop computers, but at least for now, Nvidia has little motivation to push this technology for the average consumer: “Perhaps in the future,” said Huang, “[this technology] could do physics on the PC, but this would need a Windows API.”