BENGALURU: Intel is on track in its promise to deliver 100-fold increase in performance of chip sets for artificial intelligence applications, said Naveen Rao, General Manager of Artificial Intelligence Product Group (AIPG) at Intel.
Bolstered by the acquisition of Nervana Systems, a startup developing specialised AI systems ranging from software to chips, Intel promised in late 2016 that they would be able to achieve a 100-fold increase in performance before the end of the decade. They even started integrating Nervana’s technology into its own product roadmap.
“It is quite impressive in my life time to see such a quick ramp up in performance. It (the 100-fold increase) is on track. I will give details in the next few months at the AI devcon in March,” said Naveen Rao, General Manager of AIPG at Intel.
The first version of the chip set code-named Lake Crest is designed for deep learning purposes. It is currently made available only to select set of customers like Facebook to train complex neural networks to share insights on the performance of the chips.
Early reports suggest that the Nervana Neural Network Processor (NNP) alluded to as Lake Crest has drastically reduced the time to train AI models. Intel however has not shared any insights on the performance of their NNP.
Deep-learning application that do face recognition, voice recognition, and other AI applications require huge amounts of data which perform multiple number-crunching processes called ‘data parallelism’. The NNP chips are efficient in such parallel applications. “Nervana would be install in the data center part of Intel’s business and we have multiple generations in the pipeline,” Rao said.
Complex AI applications are not suited to run on general purpose chips which Intel has been making for year. But, the chip-maker has started integrating Nervana’s IP into Intel products.
“We are actively working with all the other groups in Intel and influenced the architecture of Xeon as well as FPGAs. Obviously they are being developed in separate groups, but we have an architecture group that discusses how to incorporate NNP in theses different products and enhance it,” Rao added.