Editors’ Day Presentations

Now that we glanced at much of CUDA’s target market, we’ll quickly look into the couple of applications that were presented during Editors’ Day by guest speakers. While it might seem rather shallow to just list them one after the other, NVIDIA considered these presenters important and representative enough to showcase their work to the press. We’d be inclined to agree, and therefore why couldn’t that information also be valuable to our readers?

Tech-X’s GPU Lib: This is not an application per-se, but rather a library. As we explained previously, scientists often need to create their own small programs and the goal here is to make that easier for them. It’s basically a lightweight wrapper over CUDA that doesn’t expose any of the difficult aspects (although it might use some of them internally, we’re not sure) and uses a simplistic non-C syntax that should be very easy to understand.

It’s completely free for non-commercial purposes (including research) and NVIDIA doesn’t expect people using this to really buy Teslas for this either; so in both cases, it’s all about getting known to the world. Tech-X gets some consulting business (which is one of their main revenue streams), and NVIDIA gets more people interested in CUDA.

TechniScan’s UltraSound CT: More than 40K women will die from breast cancer in the USA this year alone, so early (and correct) diagnosis is absolutely critical. TechniScan is a company focused on the development of a new ultrasound-based diagnosis method (which was thought of decades ago in academia, but never deployed in the past) and while they’ve been able to demonstrate feasibility several years ago already, substantial development was still necessary and performance wasn’t where it had to be.

Over time, however, only the latter problem remained, and it temporarily became a stumbling block because doctors wanted the results to be ready fast enough to be able to analyze them with the patient during a single one hour visit. Normally, you’d just throw more hardware at the problem, but sadly CPU many-node scaling just wasn’t good enough to achieve the necessary performance:

Then one day one of their engineers discovered CUDA and ported the code to it (on a cheap GeForce development environment) in his spare time. Much of the algorithm was limited by FFTs, so it was relatively easy to get that part accelerated and prove it could solve their problems. While the speedup isn’t mind-blowing (we aren’t sure, but we suspect those are G80 Teslas though), the real point is that it’s fast enough for the target workload. While it might have been possible to get better scaling with additional nodes, it would likely have been very difficult and still have cost a lot more money per unit.

Using ultrasound for breast cancer detection isn’t new. What is new, however, is ultrasound CT (imaging); in traditional workflows, the technologist/radiologist must manually move a machine over the patient’s breast and try to find the potential abnormality that was indicated through mammography. If there was no prior screening, the approach would be even more difficult.

Ultrasound CT, on the other hand, delivers a full picture of the breast and is fully automated, non-invasive, high-contrast, and comfortable. They could be used for screening instead of mammography (and some studies are looking into that apparently), but it remains slower than mammography so there’s little point outside of high-risk patients. Its direct competitors are classic ultrasound and MRI, both much more expensive for a variety of reasons.

TechniScan’s technology hasn’t yet been cleared by the FDA (i.e. clinical studies are underway to prove that the solution does what they claim it does) but they currently hope to ship for revenue next year. We wanted to get an idea of how much of an opportunity this could represent for both TechniScan and NVIDIA, and they very nicely got back to us on that - cheers! :)

Apparently, they currently estimate there to be about 16,000 potential users in the USA for their solution (hospitals, women’s health centers, some general practice offices). Out of that, they expect market saturation of about 50% and further sales of ~1K units per year. If we count only the initial sales (presumably over a period of just a few years) and assume NVIDIA earns a gross profit of ~$1K per Tesla, that’s over $30M and obviously even more for TechniScan.

Given that this is just one application out of many, this is pretty impressive stuff. To put it in perspective, that’s approximately as much money as the PlayStation 3 royalties in the console’s first 9 months based on volume shipments of 6 million units. And unlike the Playstation 3 (excluding its enviable impact on the Folding@Home project), this thing is actually going to save lives too.