SAN JOSE, Calif. Advanced Micro Devices published Tuesday (Aug 14) the first of several plans to extend the x86 instruction set to ease the job of programming multicore processors. With the move, AMD is applying once again its formula of competing with Intel Corp. on technology by advancing the CPU architecture ahead of its larger archrival.
AMD announced plans for two instructions that would help programs gauge their performance in real-time. The so-called lightweight profiling extensions would give software access to information about cache misses and retired instructions so they can optimize data structures for better performance.
The move is the first of several in the works at AMD under an initiative dubbed Extensions for Software Parallelism. Other extensions may focus on hardware support for fast context switching or transactional memory, two features also seen as key for programming multicore chips
"As we look ahead in multicore systems, we see the need for hardware and software to work together to evolve a new set of techniques," said Earl Stahl, vice president of software engineering at AMD. "We see this as the first of a family of family of extensions, and we are working on others," he added.
Lightweight profiling refers to real-time analysis a program does on its performance without incurring the performance hit of going through the operating system or driver software. By delivering processor performance information in a standard way to such processes, the software could, for example, more quickly shift data structures to reduce cache misses.
AMD's proposal involves two new instructions for the x86 and one new data structure. AMD likely will also provide a software library to assist the profiling process.
"This is primarily for developers of run-time environments, certain tool sets and some complex, high performance applications such as databases," Stahl said.
Intel provides a number of compilers, software primitives and tools for its processors, increasingly aimed at enabling parallel programming techniques. In June it rolled out new compilers that can simultaneously optimize programs for vector and scalar programming.
Intel has also disclosed research in application-specific languages that could open the door to more parallel processing. However, the chip maker has not yet discussed any plans to extend the x86 instruction set for bolster parallel programming on its multicore CPUs.
For its part, AMD gained technical recognition for pioneering the transition to a 64-bit x86 architecture. However, in other areas such as media processing, Intel, not AMD, typically provides technical leadership.
Whether the new parallel programming extensions help AMD catch a meaningful edge in performance remains to be seen. AMD gave no timeframe for when it expects to end the feedback process on the new instructions, when it will put them into a CPU or how much they might boost performance.
"This will give AMD some advantage over Intel, but I doubt it would be as much as the 64-bit extensions did," said Nathan Brookwood, principal with market watcher Insight64 (Saratoga, Calif.) "It will take awhile for this to show up in processors," he added.