Linley Newsletter: November 29, 2018

 weSRCH's Best of the Internet Award
  29th-Nov-2018
 388


Linley Newsletter

Please feel free to forward this to your colleagues

Issue #628

November 29, 2018


Independent Analysis of Microprocessors and the Semiconductor Industry

Editor: Tom R. Halfhill

Contributors: Linley Gwennap, Mike Demler, Bob Wheeler


In This Issue:

- NXP's i.MX RT600 Listens for Less

- NovuMind Relieves Tensor Headaches


NXP's i.MX RT600 Listens for Less

By Bob Wheeler

NXP is cutting the cost of the consumer and IoT devices that include a voice-based user interface. Its new i.MX RT600 adds a DSP core to process audio as well as accelerate deep learning. The RT600 joins the company's other i.MX RT "crossover" products, which have many MCU features but deliver performance closer to that of application processors. They break the MCU mold by omitting embedded flash memory, thereby enabling the use of more-advanced process technologies that offer higher performance as well as lower power. The RT600 is the company's first crossover chip built in a 28nm FD-SOI process.

Whereas the i.MX RT10xx family uses a single Arm Cortex-M7 CPU for all processing, the new RT600 integrates a newer Cortex-M33 CPU and a Cadence HiFi 4 DSP operating at up to 300MHz and 600MHz, respectively. Another standout feature is 4.5MB of integrated SRAM, which both the CPU and DSP can access. The chip's security features include Arm TrustZone-M, secure boot, a physically unclonable function (PUF), and crypto engines. It has numerous serial interfaces plus a 12-bit A/D converter. The RT600 comes in a 9mm VFBGA-176 that's only 1.0mm high. It's currently sampling, with production expected in 2Q19; NXP didn't announce pricing.

For voice-based applications, the RT600 is powerful enough to perform keyword detection and other deep-learning tasks, yet it's efficient enough to serve in battery-powered devices. It competes primarily with MCUs based on Cortex-M7, such as STMicroelectronics' STM32H7-series; these alternatives are less power efficient for DSP tasks. The RT600's downsides are a more complicated programming model compared with a CPU-only design and a lack of deep-learning software. As a result, we expect mainly designers of battery-powered devices will adopt NXP's new chip.

Microprocessor Report subscribers can access the full article:

http://www.linleygroup.com/mpr/article.php?id=12064

NovuMind Relieves Tensor Headaches

By 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%.

Microprocessor Report subscribers can access the full article:

http://www.linleygroup.com/mpr/article.php?id=12065

About Linley Newsletter

Linley Newsletter is a free electronic newsletter that reports and analyzes advances in microprocessors, networking chips, and mobile-communications chips. It is published by The Linley Group. To subscribe, please visit: http://www.linleygroup.com/newsletters/newsletter_subscribe.php

Domain: Electronics
Category: Semiconductors

Recent Newsletters

Linley Newsletter: May 9, 2019

Linley Newsletter Please feel free to forward this to your colleagues Issue #651 May 9, 2019 Independent Analysis of Microprocessors and the Semiconductor Industry Edit

09 May, 2019

Linley Newsletter: April 11, 2019

Linley Newsletter Please feel free to forward this to your colleagues Issue #647 April 11, 2019 Independent Analysis of Microprocessors and the Semiconductor Industry

11 April, 2019

Linley Newsletter: April 4, 2019

Linley Newsletter Please feel free to forward this to your colleagues Issue #646 April 4, 2019 Independent Analysis of Microprocessors and the Semiconductor Industry Ed

05 April, 2019