» Current | 2019 | 2018 | 2017 | Subscribe

Linley Newsletter

Spring Hill Sets AI Efficiency Mark

September 3, 2019

Author: Linley Gwennap

Intel’s second Nervana-branded device delivers power-efficient AI-inference capability for servers. The Nervana NNP-I 1000, code-named Spring Hill, can achieve 4.8 trillion operations per second (TOPS) per watt. On ResNet-50, it offers 5x better performance per watt than Nvidia’s T4 and Xavier products, according to preliminary scores Intel showed at the recent Hot Chips conference. Built in the company’s new 10nm process, the chip combines two x86 CPUs with 12 deep-learning cores optimized for 8-bit integer (INT8) operations. Intel showed a version of Spring Hill in a tiny M.2 card that consumes just 10W, but the chip can scale to 50W to maximize performance.

The M.2 card implements an Open Compute Project (OCP) standard for low-power accelerators, offering up to 48 TOPS of deep-learning performance. Several of these 10W modules plug into Glacier Point, a PCIe extension card. Facebook leads the OCP effort, which includes the Yosemite server platform. For traditional servers, the 50W Spring Hill can reach 92 TOPS, delivering more ResNet-50 performance than Nvidia’s leading inference accelerator, the Tesla T4. Compared with the T4, Spring Hill has more on-chip memory but much lower DRAM bandwidth.

Intel plans to release the first NNP-I products by the end of this year, about the same time as the first Nervana NNP-T products. The company supports Spring Hill in its OpenVino software platform, which accepts neural networks created in TensorFlow, Caffe, MXNet, and other popular frameworks. OpenVino runs these networks on Intel CPUs, GPUs, and FPGAs as well as its Movidius accelerators, which require only a few watts. Intel’s nGraph platform also supports the NNP-I.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Events

Linley Fall Processor Conference 2019
October 23 - 24, 2019
Hyatt Regency, Santa Clara, CA
Linley Spring Processor Conference 2019
Held April 10 - 11, 2019
Proceedings available
More Events »

Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »