ElcomSoft has recently announced the switch to OpenCL, an open cross-platform architecture offering universal, future-proof accessibility to a wide range of acceleration hardware. We’re actively using GPU acceleration for breaking passwords faster. No issues with NVIDIA hardware, but working with AMD devices has always been a trouble.

So we jumped in, embedding OpenCL support into Elcomsoft Phone Password Breaker and Wireless Security Auditor. As an immediate benefit, we were able to add long-awaited support for AMD’s latest generation of graphic accelerators, the AMD Radeon™ HD 7000 Series currently including AMD Radeon™ HD 7750, 7770, 7950, and 7970 models. Headache-free support for future generations of acceleration hardware is icing on the cake.

After switching to OpenCL, we further optimized acceleration code for AMD hardware, squeezing up to 50% more speed out of the same boards. This isn’t something to sniff at, as even a few per cents of performance can save hours when breaking long, complex passwords.

OpenCL vs. CUDA

AMD goes OpenCL. What about NVIDIA? Technically, we could have handled NVIDIA accelerators the same way, via OpenCL (it’s a cross-platform architecture, remember?) In that case, we would be getting a simpler, easier to maintain product line with a single acceleration technology to support.

However, we’re not making a full commitment just yet. While some of us love open-source, publicly maintained cross-platform solutions, these are not always the best thing to do in commercial apps. And for a moment here, we’re not talking about licensing issues. Instead, we’re talking sheer speed. While OpenCL is a great platform, offering future-proof, headache-free support of future acceleration hardware, it’s still an extra abstraction layer sitting between the hardware and our code. It’s great when we’re talking AMD, a company known for a rather inconsistent developer support for its latest hardware; there’s simply no alternative. If we wanted access to their latest state-of-the-art graphic accelerators such as AMD Radeon™ HD 7000 Series boards, it was OpenCL or nothing.

We didn’t have such issues with AMD’s main competitor, NVIDIA. NVIDIA was the first player on this arena, being the first to release graphical accelerators capable of fixed-point calculations. It was also the first to offer non-gaming developers access to sheer computational power of its GPU units by releasing CUDA, an application programming interface enabling developers use its hardware in non-graphical applications. From the very beginning and up to this day, CUDA maintains universal compatibility among the many generations of NVIDIA graphical accelerators. The same simply that can’t be said about AMD.

So is it the “if it ain’t broke, don’t fix it” approach? Partly, but that’s just one side of the coin. CUDA simply offers better performance than OpenCL. The speed benefit is slight, but it is there, and it’s significant enough to get noticed. We want to squeeze every last bit of performance out of our products and computers’ hardware, and that’s the real reason we’ll be staying with CUDA for as long as it’s supported – or until OpenCL offers performance that can match that of CUDA.

Did we make the switch half-heartedly? Nope. We’re enthusiastic about the future of OpenCL, looking forward to run our software on new acceleration platforms. But we don’t want to abandon our heritage code – especially if it performs better than its replacement!