Intel Answers Our Questions, Knights Corner Demystified

Intel made the official announcement of the Knights Corner parallel processing accelerator but some things weren't perfectly clear, so we reached out to Intel, who were able, and willing, to answer a few of our questions.

To offer some context, Knights Corner is Intel's answer to the NVIDIA Tesla and all other GPU computing modules spreading across the HPC (high-performance computing) market.

It claims to have a much higher performance than the strongest such PCI Express GPU-based adapter, in fact (1 TFLOPs or more compared to 665 GFLOPS).

We covered the official bits in this article, but we felt there were some points that weren't perfectly clear or at least could use some further exploration.

To that end, we e-mailed an Intel representative with some questions.

The first thing we requested was for some clarity to be shed on the form factor of the product and its implications.

The Knights Corner, KNC for short, from the press photos is a chip instead of a PCI Express accelerator, which could mean that it might need special hardware.

As such, we asked if it will be sold as part of PCI Express adapters like GPU compute accelerators or in other guises. The company representative said, in a way, yes to both.

"No. KNC will not require any special platforms/motherboards. It will use industry standard PCI Express form factor. It will not change how supercomputers are built. The key advantages of KNC are: breakthrough performance (1TFLOPS+), easy programming model (ability to use same code and same programming techniques as for programming Xeon chips – no need to learn/use any proprietary language).

"There may be some confusion as the name Knights Corner is used both for chip and the board. And actually Knights Corner will be available as a PCI Express card however we are evaluating offering other form factors as well."

The representative elaborated on the competition with GPGPU, saying that, in addition to performance, Knights Corner has another advantage, namely familiarity.

"We believe it will be a much more competitive solution than GPGPU as it will save time and resources on not forcing people to learn any new programming models but using the code they already have and language/programming models they are most familiar with."

Now all that is left is to see how things go and if NVIDIA is quick enough to make a new and better GPU device in a timely fashion.