Hot Chips Conference: Intel releases latest details of neural network processor

At the 2019 Hot Chips Conference, Intel announced the latest details of its upcoming high-performance artificial intelligence (AI) accelerator, the Intel® Nervana™ Neural Network Processor, including the NNP-T for training and the NNP-I for inference. Intel engineers also introduced details of hybrid chip packaging technology, Intel® Optane™ data center-level persistent memory, and optical I/O chiplet technology.

 “To realize the future vision of ‘AI everywhere’, we must address the data backlog, ensure that enterprises can effectively use data, process data where it is collected when necessary, and use upstream data and computing resources more wisely. Data centers and clouds need to provide high-performance and scalable general-purpose computing and specialized acceleration for complex AI applications. In the future vision of AI everywhere, a comprehensive solution is needed from hardware to software to applications.”

——Naveen Rao, Vice President and General Manager of the Artificial Intelligence Products Group at Intel Corporation

Transforming data into information and then into knowledge requires the coordination of hardware architecture and packaging, memory, storage, and interconnect technologies, which are constantly evolving and supporting emerging and increasingly complex application scenarios and AI technologies. AI-specific accelerators can provide customers with the right intelligence at the right time, such as the newly built Intel Nervana NNP.

Intel Nervana NNP-T – Built from the ground up to train deep learning models at scale. The Intel Nervana NNP-T (neural network processor) continues to push the boundaries of deep learning training. It prioritizes two key practical factors: training the network as fast as possible and completing the training within a given power budget. This deep learning training processor is built with flexibility in mind and balances compute, communication, and memory. While Intel® Xeon® Scalable processors have added AI instruction sets and provide a good foundation for artificial intelligence, the NNP-T is built from the ground up with the features and requirements needed to handle large models without having to provide the extras needed to support traditional technologies. To meet future deep learning needs, the Intel Nervana NNP-T is flexible and programmable so it can be customized to accelerate a variety of workloads, both existing and emerging.

Intel Nervana NNP-I – High-performance deep learning inference for major data center workloads. Designed for inference and to accelerate deep learning deployments at scale, the Intel Nervana NNP-I leverages Intel’s 10nm process technology and Ice Lake cores to deliver industry-leading performance per watt across all major data center workloads. In addition, the Intel Nervana NNP-I offers a high degree of programmability without compromising performance or power efficiency. As AI becomes ubiquitous in all workloads, having a dedicated inference accelerator that is easy to program, has low latency, allows for fast code porting, and supports all major deep learning frameworks will allow businesses to leverage the full potential of data into actionable insights.

Lakefield - a hybrid core in a three-dimensional package. Lakefield is the industry's first product to adopt 3D stacking and IA hybrid computing architecture, and will be used in a new class of mobile devices. Leveraging Intel's latest 10nm process and Foveros advanced packaging technology, Lakefield has significantly reduced standby power, core area, and package height compared to previous generations of technology. The top computing performance and ultra-low thermal design power brought by Lakefield allow new slim-appearance devices, two-in-one devices, and dual-screen devices to be always online and always connected with extremely low standby power consumption.

TeraPHY - In-Package Optical I/O Chiplets for High-Bandwidth, Low-Power Communications. Intel and Ayar Labs have demonstrated the industry’s first integration of Monolithic In-Package Optics (MIPO) with a high-performance System-on-Chip (SOC). Ayar Labs’ TeraPHY* optical I/O chiplet is co-packaged with an Intel Stratix 10 FPGA using Intel’s Embedded Multi-die Interconnect Bridge (EMIB) technology to deliver high-bandwidth, low-power data communications from the chip package over distances of up to 2 km for a limited period of time. This collaboration will remove inherent bottlenecks in performance, power, and cost in data transmission, helping to build computing systems for the next phase of Moore’s Law in a new way.

Intel Optane DC persistent memory – architecture and performance. Now shipping in volume, Intel Optane DC persistent memory is the first product in the memory/storage hierarchy to be called persistent memory and is a new layer in the hierarchy. Based on Intel® 3D XPoint™ technology, Optane DC persistent memory has a memory module form factor that delivers ultra-large capacity at near-memory speeds, nanosecond low latency, and localized persistent storage. Details of the two operating modes (Memory Mode and App Direct Mode) and performance examples show how this new layer enables a complete re-architecture of the data provisioning subsystem to support faster and new workloads.