Fujitsu reported the advancement of an AI processor it alludes to as a Deep Learning Unit (DLU). Notwithstanding the organization's intends to fabricate an AI supercomputer with 24 Nvidia DGX-1 frameworks, Fujitsu likewise plans to create an AI processor conveying more than ten times the execution for each watt than the opposition by 2018.
The way that the organization is building up a DLU isn't amazing all by itself. Fujitsu has over three many years of experience creating AI and related advancements, and its K Computer sits at number eight on the best 10 rundown of supercomputers on the planet. We are, however, rather shocked the amazing 10x execution objective and, despite the fact that Fujitsu isn't naming names, everybody realizes that the "opposition" they are alluding to are industry heavyweights like Nvidia, Google, Intel, and AMD.
The greater part of the previously mentioned organizations are vigorously put resources into the AI and profound learning space. To state the opposition is high would be a huge modest representation of the truth.
Takumi Maruyama, senior chief of Fujitsu's AI Platform Business Unit, laid out Fujitsu's DLU designs at ISC 2017. Maruyama has been included in SPARC processor improvement since 1993 and is presently heading up the DLU extend. As per the graphs given by Maruyama, Fujitsu's DLU coprocessor is intended for profound learning workloads and low power utilization.
The chip highlights 16 profound getting the hang of preparing components (DPE). Each contains eight single-guideline, numerous information (SIMD) execution units with an adaptable outline using the organization's Tofu interconnect innovation.
The DLU locally bolsters FP32, FP16, INT16, and INT8 information sorts and depends vigorously on bring down exactness math to enhance both execution and vitality proficiency for handling neural systems. Fujitsu expressed that the mix of a couple of vast centers (Master) with numerous little execution centers, for example, a Deep Learning Processing Unit (DPU) will bring about higher execution with less power utilization than anything at present available.
Could Fujitsu pull it off? Everything we can state is, if the organization comes even near it's 10x objective, the AI business will get compelling intriguing in 2018.