Intel CTO desires builders to construct as soon as, run on any GPU

Over 20 years in the past, the Java programming language, initially developed by Solar Microsystems, provided builders the promise of with the ability to construct an software as soon as after which have it run on any working system.

Greg Lavender, CTO of Intel, remembers the unique promise of Java higher than most, as he spent over a decade working at Solar. As an alternative of needing to construct purposes for various {hardware} and working programs, the promise of Java was extra uniform and streamlined improvement. 

The flexibility to construct as soon as and run wherever, nonetheless, isn’t uniform throughout the computing panorama in 2022. It’s a state of affairs that Intel is trying to assist change, no less than in terms of accelerated computing and using GPUs.

The necessity for a uniform, Java-like language for GPUs

“As we speak within the accelerated computing and GPU world, you need to use CUDA after which you’ll be able to solely run on an Nvidia GPU, or you’ll be able to go use AMD’s CUDA equal operating on an AMD GPU,” Lavender advised VentureBeat. “You may’t use CUDA to program an Intel GPU, so what do you utilize?”

That’s the place Intel is contributing closely to the open-source SYCL specification (SYCL is pronounced like “sickle”) that goals to do for GPU and accelerated computing what Java did many years in the past for software improvement. Intel’s funding in SYCL isn’t completely selfless and isn’t nearly supporting an open-source effort; it’s additionally about serving to to steer extra improvement towards its not too long ago launched client and information heart GPUs.

SYCL is an method for information parallel programming within the C++ language and, in response to Lavender, it seems to be rather a lot like CUDA.

Intel helps standardization for one code to rule all of them

To this point, SYCL improvement has been managed by the Khronos Group, which is a multi-stakeholder group that’s serving to to construct out requirements for parallel computing, digital actuality and 3D graphics. On June 1, Intel acquired Scottish improvement agency Codeplay Software program, which is likely one of the main contributors to the SYCL specification.

“We must always have an open programming language with extensions to C++ which are being standardized, that may run on Intel, AMD and Nvidia GPUs with out altering your code,” Lavender mentioned.

Automated instrument for changing CUDA into SYCL

Lavender can be a realist and he is aware of that there’s a lot of code already written particularly for CUDA. That’s why Intel builders constructed an open-source instrument referred to as SYCLomatic, which goals emigrate CUDA code into SYCL. Lavender claimed that SYCLomatic at this time has protection for roughly 95% of all of the performance that’s current in CUDA. He famous that the 5% SYCLomatic doesn’t cowl are capabilities which are particular to Nvidia {hardware}.

With SYCL, Lavender mentioned that there are code libraries that builders can use which are machine unbiased. The best way that works is code is written by a developer as soon as, after which SYCL can compile the code to work with no matter structure is required, be it for an Nvidia, AMD or Intel GPU.

Trying ahead, Lavender mentioned that he’s hopeful that SYCL can develop into a Linux Basis undertaking, to additional allow participation and progress of the open-source effort. Intel and Nvidia are each members of the Linux Basis supporting a number of efforts. Among the many tasks the place Intel and Nvidia are each members at this time is the Open Programmable Infrastructure (OPI) undertaking, which is all about offering an open normal for infrastructure programming models (IPUs) and information processing models (DPUs).

“We must always have write as soon as, run all over the place for accelerated computing, after which let the market resolve which GPU they need to use, and degree the enjoying subject,” Lavender mentioned.