» Current | 2018 | 2017 | 2016 | Subscribe

Linley Newsletter

NovuMind Relieves Tensor Headaches

November 27, 2018

Author: Mike Demler

Startup NovuMind has received working silicon for its first neural-network-inference ASIC, which it calls NovuTensor. Although the chip’s internal compute engines run at just 400MHz, it executes up to 15 trillion operations per second (TOPS). NovuTensor accomplishes that feat through a combination of hardware parallelism and time-division multiplexing (TDM). It streams in data at six times the core frequency, but it achieves high performance by multiplexing the data to parallel compute-tree branches running at the lower frequency. NovuMind manufactures the ASIC in GlobalFoundries’ 28nm process, and it plans to deliver production volumes in 1Q19.

The company describes its chip as a “na­tive” tensor processor owing to its matrix-multiplication method that avoids explicitly un­folding three- and higher-dimensional ten­sors. Be­cause multiply-accumulate (MAC) op­er­a­tions can account for as much as 90% of con­volutional-neural-network (CNN) op­era­tions, NovuMind de­ter­mined that elim­inating the tensor-unfolding over­head of con­ventional methods would boost MAC utiliza­tion. In an­other de­par­ture from con­vention, where­as most neural-network pro­cessors require large inte­grat­ed SRAMs to hold the huge vol­ume of ac­tiva­tion and weight data that infer­ence en­gines re­quire, NovuTensor in­tegrates only a small scratch­pad memory.

NovuMind compared NovuTensor and a Jetson Xavier development kit using just that SoC’s embedded Volta GPU, which Nvidia rates at 20 TOPS. Although NovuTensor beat Xavier on four of the six tests, it trailed by nearly 2:1 on Yolo v2 and roughly 3:1 on ResNet-50. But on ResNet-70, it jumps to 420 IPS, surpassing Xavier’s ResNet-50 performance by roughly 50%.

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 2018
Covers processors and IP cores used in embedded, communications, automotive, IoT, and server designs.
October 31 - November 1, 2018
Hyatt Regency, Santa Clara, CA
More Events »

Newsletter

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