Intel Focuses on Scale Out AI Training With New Chip

Carey Kloss has been intimately involved with the rise of AI hardware over the last several years, most notably with his work building the first Nervana compute engine, which Intel captured and is rolling into two separate products: one chip for training, another for inference. He tells The Next Platform that the real trick is to keep pace with the increasing size and complexity of training models with balanced architectures. Given that the compute required for training is doubling almost quarterly, this is more important that ever from performance, efficiency, and scalability standpoints. Kloss and Intel think they have finally struck that golden grail of balance with the Spring Crest Deep Learning Accelerator or, more simply, the Intel Nervana NNP-T. The name may lack the poetic ring of say, a "Volta" but we do see this training chip from its current status (they just got their first silicon back) as competitive and filling in some gaps in terms of performance/efficiency and data movement potential.