Graphics start-up Otoy has claimed it has developed a means of running software written for Nvidia's Compute Unified Device Architecture (CUDA) platform, allowing CUDA code to run unmodified on rival graphics cards and CPUs, in particular those from AMD.
Nvidia's CUDA is a jewel in the company's crown. Created to allow easy offloading of highly-parallel tasks to the graphics processor, CUDA is capable of accelerating workloads considerably compared to running the same tasks on a CPU - the entire premise of the company's popular range of Tesla accelerator boards. CUDA has one major drawback, though: vendor lock-in. When developing a GPGPU-accelerated application, the use of CUDA with all its high-quality documentation and resources means that your program will only ever run on Nvidia's GPUs. While more open standards like OpenCL allow code to run on almost any hardware, the support that Nvidia gives to CUDA means that it is the first choice of many development houses.
Start-up Otoy claims it has broken the lock-in cycle, however. In an interview with VentureBeat
, Otoy chief executive Jules Urbach revealed a CUDA cross-compiler capable of running unmodified CUDA code on any GPU or CPU which he claims his company created in just nine weeks.
There's a catch, of course: use of the compiler requires access to the original source code of the CUDA application in question, meaning it's not something that end-users can apply to their systems in order to start running CUDA-accelerated applications on their CPU or third-party GPU. It's also proprietary: Otoy plans to release it with version 3.1 of its Octane rendering engine and has not indicated that it will be available separately. For those who do pick it up, though, Otoy is claiming impressive results, with Urbach stating that any given application 'runs on the other cards at the same speed as it runs on Nvidia cards.
Nvidia has not commented on Otoy's claims.