Nvidia Tesla Powers An Amazing Ultrasound Scanner


Santa Clara (CA) - A few weeks ago, we learned about a technology that was fascinating simply because of the fact how current technology advances have an impact not just on how we use computers, but how they improve our life. In this case it was TechniScan Medical Systems’ GPU-assisted UltraSoundCT scanner, which may replace mammograms one day and can provide much more detailed breast scans than traditional mammograms already today. We came across a few more details how the technology involved and decided to share those with our readers.

TechniScan is a small company that started developing ultra-sound scanner systems in the late 1990s and employs about 15 people today. Most recently, the company introduced a GPU-assisted product called UltraSoundCT, which uses ultrasound to scan a breast. For a patient, this approach is much more comfortable than a mammogram, since there is no ionization radiation and no need for the compression of a breast.

Back when TechniScan began developing scanners, the main problem of the approach was that enormous computing power is needed to render scans in a timely fashion. At the same time, the volume of scans has been increasing dramatically and with obesity being an increasing problem in many societies, some believe that breast cancer could start developing among the male population as well. So far, there are only isolated medical cases, we were told, but if current rates of obesity persist, we might have influx of male patients.

According to TechniScan, doctors typically expect scan results within 15-30 minutes so that they can establish a diagnosis and prescribe the necessary treatment. Five years ago, TechniScan built a 12-node cluster powered by Pentium M processors. Scanned results were available after 118 minutes. Using a 6-node cluster, the time increased to only 135 minutes, revealing that CPU scaling would not be the answer to the problem.

The company continued to run tests with Pentium 4 and Core 2 generations of processors, but even with the fastest Core 2 Duo and Quad processors, the render time could not be cut under 45 minutes, we were told. A possible solution popped up when TechniScan senior software engineer Jim Hardwick bought a GeForce 8-series card and discovered Nvidia’s CUDA SDK. Jim is an avid gamer, so he bought the card to enjoy latest and upcoming games, but a quick run of his code on a GPU apparently lit up more than just one bulb. Fast forward to 2008 - today the code is ported to CUDA and utilizes four Tesla C870 boards. The render time was cut from 45 minutes on a 16-core Core 2 cluster to only 16 minutes.

In TechniScan’s case the most time is consumed by data preparation since the system has to process large datasets and produce images that doctors can rely on. The GPU scored major speedups in calculating 2D FFT’s, in which a single 8800 GTX was eight times faster than a Core 2 Quad at 2.66 GHz, while complex exponentiation with 12 million elements ended up being accelerated by a factor of 320x. Complex Exponentiation is usually run 50-60 times, so you can see how a 4-GPU setup was able to cut the total rendering time.

The performance increase was not the only benefit the GPU technology delivered. The power of a GPU is packaged in a relatively small product that has a low total cost of ownership. The price of a complete system is critical in many scenarios. But in this case, it is not an important part of the equation. A four-GPU setup checks in at about $15,000 - $20.000 - just a fraction of the complete system, which is priced at about $1 million.

Hardwick pointed out that this system is not a replacement for traditional mammograms, but rather an affordable tool that can be used in early detection and prevention. TechniScan is expecting a lot from next-generation Tesla products, which are set to offer even more horsepower to render the results even faster - possibly in the 10 minute range. Currently, the UltraSoundCT is in the middle of clinical trials as a part of two-step process in getting approval from the FDA. If everything goes according to plan, we should see this system coming to market within the next two years.