The GPU That Never Was
Intel’s canceled project to create a GPGPU chip—a GPU that could be used for CPUlike tasks—is interesting, because it foreshadowed today’s multicore world.
Back in 2008, when Intel was pushing Core 2 Duo as its mainstream desktop solution, the company announced it was going to make a graphics card that was a little bit different. It was based on the original Pentium, and eschewed modern trappings such as out-of-order execution and multiple instructions per clock. What it could do, however, was squeeze 10 of its cores into a space no bigger than the Core 2 Duo’s two cores.
Graphics cards are built from streaming processors, and Intel’s Larrabee lagged behind the core counts of other GPUs of the time, but it had one distinct advantage: While Nvidia’s cores could only work on one operation at once, and AMD’s could manage five, Intel’s could manage 16 operations. This meant it needed a lot of help from the CPU to maximize the utilization of its resources, but Intel had this covered in its CPU designs. The other thing that set Larrabee apart was its programmability—it was completely programmable, using the X86 architecture programmers were already familiar with, while the more dedicated GPUs were only partly programmable at best.
Unfortunately, Larrabee, as initially envisaged, never appeared. Development of the chips fell behind schedule, and the competitiveness of the GPU market left it behind, so Intel canned it. It lived on, however, as Xeon Phi, a series of co-processor boards developed for supercomputing, with up to 72 cores running at up to 1.7GHz. The Tianhe-2 supercomputer in Guangzhou, China, is made up of 16,000 nodes, each of which comprises two Ivy Bridge Xeon CPUs and three Xeon Phi co-processors, for a total of 3,120,000 cores. Each node has 88GB of RAM. The Tianhe-2 was the fastest computer in the world from June 2013 to June 2016.
Intel will ship Xeon Phi chips until July 2019, but after that they’re gone forever. Interestingly, Tom Forsyth, the chief architect of Larrabee, is back at Intel as part of the team headed by former AMD graphics boss Raja Koduri, thought to be preparing Intel’s entry into the discrete GPU space.