The first of February in Munich launched European Conference AI-Day. A similar event was held in the US in the 20 November numbers.
The ownership of Intel company Nervana passed in August last year. It is assumed that in the first half of this year, Intel will begin to release engineering models of innovative accelerators Lake Crest. Created solutions Intel promises will soon lead to a deep machine learning accelerators, which accelerate hundredfold perform specific tasks related to pattern recognition, speech and so on. This should happen by 2020.
The first, as already mentioned, boosters Lake Crest . Accelerators will be held on 12 computing clusters interconnected by a new high-speed bidirectional bus, each channel which provides bandwidth up to 100 Gbit / s. Note that the company uses the term "cluster" instead of cores, which involves splitting each cluster into a plurality of ALU or other specific units. In fact, Intel continues the work of GPGPU, but instead initially shader processors, as did the NVIDIA and AMD,
advances in the field of machine learning some deep computing cores.
It is said that Lake Crest is based on a different microarchitecture - namely tensor. The scalar or vector architecture, typical of classic GPU, Intel has decided not to use. Instead, it will operate with tensors that represent data as multidimensional arrays. Due to this, Intel on level ground expects to see a 10-krtanoe advantage of new solutions in the field of deep machine learning.
At one site with "CPU" Lake Crest will be located up to four stacks HBM2 capacity up to 8 GB each. Total processor will be targeted to the 32 GB HBM2 memory with a total capacity of 1TB / s. Note for assembly Xeon Phi accelerator, Intel uses Micron HMC memory (but calls it MCDRAM). Definitely on HBM2 chip has a competitor.
Subsequently, Intel plans to release a hybrid processor Knights Crest. In this solution a single substrate will combine x86-compatible core (Xeon), Lake Crest unit and possibly, FPGA-matrix. Related Products :
|