Linley Newsletter: January 5, 2017

 weSRCH's Best of the Internet Award

Linley Newsletter (Formerly Processor Watch, Linley Wire, and Linley on Mobile)
Issue #529
January 5, 2017

Independent Analysis of Microprocessors and the Semiconductor Industry

Editor: Tom R. Halfhill
Contributors: Loyd Case, Linley Gwennap, David Kanter, Loring Wirbel

In This Issue:

- Year in Review: Mobile Processor Market Rebounds
- Macom to Grab AppliedMicro
- Cortus APS29 Doubles Instruction Fun
- Everspin MRAM Targets Enterprise
- Many Options for Machine Learning

Available Now:
"A Guide to Processors for Wireless Base Stations" (4th Edition)

By Jag Bolaria and Tom R. Halfhill

The Linley Group is pleased to announce the release of "A Guide to Processors for Wireless Base Stations" (4th Edition). This new edition begins with an overview of 3G (UMTS) and 4G (LTE) technology, highlighting various aspects of the cellular network, underlying technology, and relevant semiconductor technologies that affect the design and capabilities of integrated base-station processors.

Another chapter discusses major technology and business trends, providing insight into how this market segment will evolve. We also examine new 5G technologies and provide detailed market projections for residential, enterprise, and macro base stations. We cover shipments of integrated base-station processors by each vendor and forecast the potential revenue by market segment out to 2020.

Following these introductory chapters, the report covers the major suppliers of integrated base-station processors: Broadcom, Cavium, Intel, NXP (which acquired Freescale), Qualcomm, and Texas Instruments. We provide a business synopsis, explain their products' key features, and draw conclusions about their competitive standing and the system designs for which their offerings are best suited. We also discuss the internal architecture of their processors, system-design considerations, and their likely product roadmaps.

We then dedicate a chapter to detailed head-to-head comparisons of many of the available base-station processors. The report concludes with our outlook for the leading vendors.

A detailed description and the table of contents can be found on our website:

...or order the report by emailing us at

Year in Review: Mobile Processor Market Rebounds
By Linley Gwennap

After a stomach-churning 13% drop in 2015, cellular-processor revenue should rise 5% in 2016, recovering some of the loss. Although the smartphone market remains weak, cellular tablets and cellular IoT provide some bright spots for growth. LTE is already the dominant protocol as operators begin to shut down their 2G and CDMA networks.

Following Intel's withdrawal, the smartphone market is now supplied almost entirely by three merchant processor vendors and three large OEMs that design their own mobile processors. Among the merchant vendors, MediaTek is rapidly gaining share, while Qualcomm is declining. Samsung adopted Qualcomm processors for some of its Galaxy S7 models (after completely shunning the Snapdragon chips for the previous generation), but the largest smartphone maker is moving more of its lower-cost models to in-house Exynos processors and to MediaTek, reducing opportunities for Qualcomm.

Qualcomm continues to lead at the high end, where its Snapdragon 820 won most flagship Android designs. After the Snapdragon 810 stumbled, the 820 provided better CPU efficiency as well as industry-leading graphics and LTE performance. Samsung delivered its first flagship processor with an integrated LTE modem and its first in-house CPU design, approaching Qualcomm's level of innovation. Huawei continues to offer strong products in its Kirin line. But Apple's A10 processor crushed the competition in single-core performance and added a new "little" core for greater power efficiency.

Shipments of wearables have yet to take off, but processor vendors are increasing their investing in this smaller market. In the past year, Qualcomm created the Snapdragon Wear line and MediaTek extended its broad portfolio. Samsung and Apple also deployed multiple new custom processors in 2016, which appear in products such as the Gear S3 smartwatch and Apple's Airpods.

Microprocessor Report subscribers can access the full article:

Macom to Grab AppliedMicro
By Loring Wirbel

Hoping to increase its data-center presence, high-speed-analog specialist Macom has offered $776 million for AppliedMicro, a developer of communications and processor products. AppliedMicro's advanced DSP talents in PHY modulation could augment Macom's expansion in silicon photonics for modules that serve 100Gbps and 400Gbps data-center applications. The majority of AppliedMicro's revenue, however, derives from optical transport network (OTN) mappers.

Macom clarified that it's only interested in AppliedMicro's connectivity business and that it will seek to divest the computing business, which includes legacy embedded processors and ARMv8 server processors. The divested elements could include PowerPC and X-Gene products, depending on buyer interest. Macom anticipates completing the divestments within 100 days of closing the acquisition.

The acquisition will end an attempted turnaround by AppliedMicro, which has struggled to sustain growth over the past decade. In recent years, the company has used its profitable communications business to fund development of its X-Gene server processors, which have yet to generate meaningful revenue. Meanwhile, sales of PowerPC embedded processors have begun a terminal decline, leading AppliedMicro's overall sales to drop. The size of the acquisition offer indicates that Macom believes it can sell the struggling computing business, or at least X-Gene, for a substantial sum.

AppliedMicro has had a healthy optical- and TDM-interface business for more than 20 years, since its historical involvement in Sonet framers and mappers (when it was known as Applied Micro Circuits Corp, or AMCC). In recent years, its X-Weave high-speed-interface business kept a low profile, as the company emphasized X-Gene. Connectivity products, however, accounted for 72% of the revenue in AppliedMicro's most recent fiscal quarter.

Microprocessor Report subscribers can access the full article:

Cortus APS29 Doubles Instruction Fun
By Loyd Case

Cortus developed the new APS29 dual-issue CPU IP (intellectual property) for customers seeking better single-core performance. Based on the Cortus v2 instruction set, the APS29 boosts performance without adding the complexity of multicore programming. It has two multipliers and enables 64-bit buses, a first for the company. Customers can license the APS29 RTL now. The main applications include IoT and embedded security.

The new CPU targets users of the earlier APS25 who would have needed to use the processor's dual-core mode to improve performance. That older model includes hooks for multicore configurations and supports data-cache coherence for customers desiring a single, shared L1 cache. Each APS25 CPU would still need individual instruction caches, however.

Owing to the dual-instruction decoders, the APS29 consumes 2.26x more die area than the APS25 at 0.233mm^2 in a 90nm process. Maximum clock frequencies differ as well: 425MHz for the APS25 and 360MHz for the APS29. Despite the slower clock, the APS29 outperforms its predecessor, generating 1,325 CoreMarks versus 1,071 for the APS25. The new core uses 0.039mW/MHz, double the power per megahertz of the APS25. Cortus estimates it can clock at up to 1.4GHz if built using a 28nm TSMC process.

Cortus logged 900 million processors shipped in 2015 using its cores, and it's on track for 1.2 billion in 2016, a 33% year-over-year increase. Its main competitor, Andes Technology, saw a 58% increase in 2015, but its growth has stalled in 2016. The APS29 gives Cortus a higher-performance product to keep customers from defecting, and it should help the French company keep its revenue stream flowing.

Microprocessor Report subscribers can access the full article:

Everspin MRAM Targets Enterprise
By David Kanter

Having shipped 60 million units, Everspin is a pioneer in magnetoresistive RAMs (MRAMs): nonvolatile CMOS-compatible storage that offers DRAM-like performance as well as greater endurance and retention than flash memory. The company has developed three types of bit cells, employing them in several product generations.

Compared with flash memory, MRAM uses a fundamentally different data-storage mechanism that relies on changing the cell resistance by electrically manipulating the polarity of magnetic materials. It employs lower, CMOS-compatible voltages for programming and can be formed in the metal layers of a logic process. MRAM has high thermal and radiation tolerance and is much easier than flash to integrate in a system or chip owing to the lower voltages and wider tolerances.

Everspin has already shipped a first-generation toggle MRAM for embedded systems and a second-generation spin-torque RAM (ST-MRAM) that's compatible with DDR3 interfaces and primarily aims at storage applications (e.g., SSDs and RAID arrays). Looking forward, the company is primarily focusing on enterprise storage, where high-performance nonvolatile memories can reduce system cost. For the third-generation perpendicular ST-MRAM, the company is working with GlobalFoundries to manufacture discrete products. The two are also taking advantage of the CMOS compatibility to offer embedded ST-MRAM in a 22nm FD-SOI process that should reach production by 2018.

Microprocessor Report subscribers can access the full article:

Many Options for Machine Learning
By David Kanter

From Siri to self-driving cars, machine learning and neural networks are rapidly growing in popularity, and engineers are using everything from CPUs to specialized hardware to keep pace. Analyzing data is increasingly essential to modern companies and the economy as a whole. For example, Facebook and Google derive the vast majority of their revenue from advertising that is primarily enabled by advanced analytics. As data suffuses and enriches ordinary products and services through the Internet of Things (IoT) and automation, the need for analytics is expanding quickly.

The popularity of neural networks and machine learning in the last five years owes directly to larger training data sets and greater computational resources. This success only breeds greater adoption, compounding the need for more performance and more performance per watt. Neural-network training and inferencing are intrinsically related, but their goals and requirements are very different, leading to different hardware solutions.

For large neural networks, training is a high-performance-computing (HPC) application that requires maximum throughput and favors large, power-hungry chips; smaller training problems, however, are similar to workstation applications. Today, training commonly takes place on GPUs and highly parallel CPUs, but many companies are investing in dedicated hardware to further boost performance and power efficiency.

By contrast, inferencing is pervasive, serves in everything from email to banking, and is often latency sensitive. It can be deployed in the data center or directly on client devices. Because it has lower performance requirements than training, inferencing uses a much more diverse array of hardware solutions across the programmability spectrum.

Machine learning is still in the early stages of adoption, and software is rapidly evolving. System designers face a bewildering array of options including CPUs, GPUs, FPGAs, DSPs, and dedicated hardware (ASICs). Ultimately, their choices depend as much on the overall architecture as on the chips themselves.

Microprocessor Report subscribers can access the full article:

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 and consolidates our previous electronic newsletters: Processor Watch, Linley Wire, and Linley on Mobile. To subscribe, please visit:

Our previous subscription newsletters -- Microprocessor Report, Networking Report, and Mobile Chip Report -- are now consolidated in a single subscription newsletter: Microprocessor Report. MPR publishes articles with significantly more detail on the subjects covered in our free electronic newsletters. For subscription information, please contact us at 408.270.3772;

Back issues:

View our privacy policy here:

Domain: Electronics
Category: Semiconductors
weSRCH App on Apple

Recent Newsletters

Linley Newsletter: August 8, 2019

Linley Newsletter Please feel free to forward this to your colleagues Issue #664 August 8, 2019 Independent Analysis of Microprocessors and the Semiconductor Industry E

08 August, 2019

Linley Newsletter: August 1, 2019

Linley Newsletter Please feel free to forward this to your colleagues Issue #663 August 1, 2019 Independent Analysis of Microprocessors and the Semiconductor Industry E

01 August, 2019

Linley Newsletter: July 25, 2019

Linley Newsletter Please feel free to forward this to your colleagues Issue #662 July 25, 2019 Independent Analysis of Microprocessors and the Semiconductor Industry

25 July, 2019