Xeon Phi: an x86 coprocessor, but what about OpenCL?

By , the in Hardware - No comment
A few weeks ago, Intel restarted the hype about its abortive "Larrabee" architecture. This time the project is proceeding to the market, as proved by Intel assigning it a trade name: "Xeon Phi".

If you followed the announcement you will be aware that the new chip contains "more than 50 cores" capable of running x86 code, and up to 8GB of memory, all mounted on a PCI-express card rather than the more traditional CPU socket, to negate the need for a new design of motherboard.

Clearly this has changed the aim of the project from a GPU that is capable of GPGPU tasks, to that of a coprocessor, similar to the x87 chips of old.

The card is designed to work alone, or in a cluster as part of a wider supercomputer arrangement.

In terms of software support, details are currently unclear. The card will run a Linux OS, isolating itself completely from the rest of the system (one could ask how it functions as a coprocessor given this restriction...) and Intel's own marketing documents do not mention OpenCL. Is Intel ignoring this standard-in-the-making?

A few days ago, the Parallelis website scrutinized the documentation for the card and raised some questions about its effectiveness, particularly when running x86 code, and the limitations imposed by using PCI-express (which is much slower than interconnects between CPUs on the same chip, or CPUs and RAM, for example).

In short: this is a potentially fascinating development, but there are unanswered questions. Intel, the ball's in your court...

Source: PCInpact (in French)