Intel® Nervana™ Neural Network Processors

See the design philosophy and research behind the Intel® Nervana™ Neural Network Processors,
designed from the ground up for deep learning training and inference at massive scale.

Intel® Nervana™ NNP Architecture Revealed at Hot Chips 2019

Intel® Nervana™ Neural Network Processors for Training (NNP-T)

Intel® Nervana™ Neural Network Processors for Training (NNP-T)

To quickly process vast, sparse, or complex data for large models within a power budget, AI hardware must deliver a critical balance of compute, communication, and memory. The Intel® Nervana™ Neural Network Processor for Training (Intel® Nervana™ NNP-T) does just that. With an all-new architecture that maximizes the re-use of on-die data, the NNP-T was purpose-built to train complex deep learning models at massive scale, and simplify distributed training with out-of-the-box scale-out support.

  • Neural Network specialized Tensor Processing Clusters (TPC) provides high utilization of underlying compute
  • High-efficiency memory architecture with independent HBM and TPC-to-TPC data buses enables high efficiency scaling of complex training models
  • Glue-less Connectivity with massive scaling via Intra and Inter Chassis Links enabling cross-chassis scale-out with the same network connectivity
  • Open Accelerator Module built to Open Compute Project (OCP) specification making it OCP Accelerator Infrastructure ready

Intel® Nervana™ Neural Network Processors for Inference (NNP-I)

Intel® Nervana™ Neural Network Processors for Inference (NNP-I)

Enterprise-scale AI deployments are significantly increasing the volume of inference cycles, while demanding ever-stricter latency requirements. The Intel® Nervana™ Neural Network Processor for Inference (Intel® Nervana™ NNP-I) was built for this intensive, near-real-time, high-volume compute. By combining a CPU core and purpose-built AI inferencing engine, NNP-I delivers the novel hardware architecture that emerging, increasingly complex use cases demand, turning customer data into knowledge with an incredibly efficient, mutli-modal inferencing solution.

  • Inference compute engines provide maximum inference performance and efficiency
  • Dynamic power management and fully integrated voltage regulator (FIVR) technology optimizes SoC performance at different power envelopes
  • On-die SRAM and fabric deliver high performance for deep learning models
  • Comprehensive, standards-based deep learning software support: Framework Integration, ONNX, nGraph, OpenVINO, C++
  • Intel’s latest 10nm process technology delivers greater performance with lower power
  • Extreme programmability with on die Intel Architecture cores, including AVX and Vector Neural Network Instructions (VNNI)