In the current year will be accelerators computing Intel Lake Crest based on tensor architecture
It is no secret that some calculations are much better suited to modern GPU than the CPU. This leads to the fact that graphics cards are used to build supercomputers.
Intel also offers an alternative to the Xeon Phi accelerators. And soon to be added brand new solutions that have so far appeared under the code name of Lake Crest. The compute accelerators Lake Crest are part of the platform Nervana, which is aimed at building systems for deep learning, neural networks and so on.
About Lake Crest before information did not exist, but now the processor giant has shared some details. Data enhancers which will be performed in modules Multi-Chip Module (MCM), will act as coprocessors. In them will be based on a new architecture Flexpoint. Intel promises a performance increase in the parallel arithmetic calculations up to 10 times compared to existing solutions.
It should also be noted that the architecture Flexpoint tensor. We are accustomed to a vector or a scalar (and superscalar) architectures, GPUs, which operate the corresponding values. Ansorena architecture deals with objects, linearly transforming the elements of one linear space into elements of another.
Chip Lake Crest will contain 12 clusters, which in turn will consist of several cores. Also on the same substrate will be located four stack memory HBM2 with a total capacity of 32 GB, a frequency of 2 GHz and bandwidth of 1 TB/s.
Test samples of the new accelerators will be available for some of the Intel in the second half of this year.