The Shocking Rise of the Graphics Machines!

The popularity of graphics cards with their immensely powerful GPUs (Graphic Processing Units) are causing the HPC world to look at graphics card as the next big development. Today, NVIDIA announced a new product line called Tesla a GPU aimed squarely at the HPC market. Read on to find out more about Tesla line of products (with pictures).

Tesla to the Rescue!

NVIDIA has just announced a new GPGPU (General Purpose Graphics Processing Unit) called Tesla. The Tesla includes a somewhat new processor called the C870. Also announced were a GPU Computing server called the S870, and a desk-side unit called the D870

The new processor is called the C870 and is rumored to be a heavily modified GeForce 8800GTX card. Here are the basic stats on it:

• 1.5 GB of GDDR3 memory
• Memory bandwidth: 76.8 GB/s (384-bit bus)
• Peak performance of 518 GFLOPS (Billion floating-point operations per second)
• IEEE 754 Single precision
• 170W power consumption
• PCI Express x16 connector (takes two slots of space)
• 128 multi-threaded processors in a single GPU
• Has no video ports
• Approximately $1,499
• Requires 2 internal PCI-e connectors

Figure One: The Tesla HPC/Video Card

NVIDIA also announced two other products in the Tesla family. These product are just not graphics cards, but basically co-processors for HPC nodes. NVIDIA introduced (Figure Two) the D870 which is a desk-side unit that has up two C870 GPUs (1.0 TFLOPS of peak performance).

Figure Two: Deskside with two C870 GPUs

You can connect the D870 to another system by using an NVIDIA PCI x8 or x16 adapter card in the host node (this only adds about 10W of power usage to the host node). Then it is connected to the D870 with a PCI-e cable. There is a switch inside that box that allows the data to be sent to either of the GPUs. Note that CUDA, NVIDIA's GPGPU programming tools allow you to program or use both GPUs transparently. With 2 GPUs, the D870 has a total of 3 GB of memory that can be used. In addition, it is fairly quiet, producing only about 40 dB of noise (normal conversation can happen up to about 45-50 dBs). It produces about 550W and has a list price of about $7,500. D870 Deskside Supercomputer

The last product (Figure Three) that NVIDIA announced is the S870 which is similar to the D870, but is designed to be attached to a rack-mounted node with up to 4 GPUs (4+ TFLOPS peak performance).

Figure Three: Rack-mount with four GPUs

The S870 is a 1U rack-mount unit designed for normal 19" racks. It connects to a host node in the same way that the D870 does. In the standard configuration, the host node has 1 PCI-e adapter card driving 4 GPUs and there is an optional configuration of 2 PCI-e adapters in the host node each driving 2 GPUs. There is also likely to be a future model with 8 GPUs. The unit uses up to 800W at peak and will cost about $12,000. It would only take 1 or 2 of these units to make the Top500 (assuming peak performance).

Why Is This Announcement Significant?

Tesla was a bit of a maverick during his time. His ideas and claims were considered to be a bit outrageous leading to him being called a mad scientist, but in many ways he was a genius. His Power transmission without wires development was demonstrated as early as 1891 and only recently been duplicated. So Tesla is an appropriate title for the project - new concepts that could lead to a transformation in HPC.

This new product could be a real disruptive influence on HPC because of the huge amount of computational power in a small, relatively low-power, processor. There are already some developments where codes have been ported to GPUs with great success including one example with saw about a 240X improvement performance. The NVIDIA CUDA environment also allows you to develop using C and C++ tools and compilers that many coders are used to. More over an important development is that the C870 is IEEE 754 compliant. Other GPUs aren't really IEEE 754 compliant and this has caused problems with code development.

Equally significant is that both NVIDIA and AMD that make GPUs and graphics cards are developing product for the HPC market. The HPC market is not nearly as big as the graphics market so it is significant that these companies have developed products for the HPC market. Granted the are derivative products, but the fact that they consider the HPC market worth chasing is note worthy. Equally noticeable is that the GPUs are significantly faster than other co-processor options for HPC (such as FPGA and Clearspeed). Yes, Tesla is currently single precision, but a number of applications can take advantage of it. But, the rumor mill says that NVIDIA should be coming out with a double-precision GPU by the end of the year.

Finally, it should be noted that such astounding performance is not available to every applications. Those that map well into the GPU architecture should derive great benefit from these types of applications.

Dr. Jeff Layton hopes to someday have a 20 TB file system or a 20 TFLOP system in his home computer (donations gladly accepted). It will be interesting to see which one will happen first. He can sometimes be found lounging at a nearby Fry's, dreaming of hardware and drinking coffee (but never during working hours).