Linley Newsletter: January 10, 2018

 weSRCH's Best of the Internet Award

Linley Newsletter

Please feel free to forward this to your colleagues

Issue #634

January 10, 2019

Independent Analysis of Microprocessors and the Semiconductor Industry

Editor: Tom R. Halfhill

Contributors: Linley Gwennap, Mike Demler, Bob Wheeler

In This Issue:

- New Deep Learning Report

- Qualcomm Gets Serious About PCs

- Barefoot Joins 400GbE-Switch Club

- Year in Review: IP Suppliers Push AI Accelerators

“A Guide to Processors for Deep Learning” has been completely revised to reflect the growth in this rapidly evolving market. AI has been dominating the news with dozens of new products and technologies announced in just the last year. Deep learning isn’t limited to just data centers and self-driving cars, it is even moving into client devices such as smartphones and embedded (IoT) systems.

This completely revised report focuses on the hardware technologies and products driving this market. You’ll find detailed technical coverage of announced chip products from AMD, Eta Compute, Graphcore, Gyrfalcon, Habana, Intel (including former Altera, Mobileye, Movidius, and Nervana technologies), Mythic, NXP, Nvidia (including Tegra and Tesla), Qualcomm, Wave Computing, and Xilinx. It also covers IP cores from AImotive, Arm, Cadence, Cambricon, Ceva, Imagination, Synopsys, Videantis, and the open-source NVDLA. Other chapters cover Google’s TPU family of ASICs and Microsoft’s Brainwave.

Order by January 31 to take advantage of the introductory offer. For more information on this new edition, visit our website.

Qualcomm Gets Serious About PCs

By Linley Gwennap

The supercharged Snapdragon 8cx is Qualcomm's first PC processor that features a unique silicon design, emphasizing the company's growing investment in and commitment to mobile PCs. The chip reuses many elements of the new Snapdragon 855 smartphone platform but boosts CPU and particularly GPU performance to better suit the needs of larger devices.

To save power and die area in the 855, just one of the four Cortex-A76 CPUs, designated the "prime" core, is optimized for maximum clock speed and performance. The Snapdragon 8cx instead features four prime cores, enhancing performance. The 8cx also includes four low-power Cortex-A55 CPUs to increase efficiency on less demanding tasks. It doubles the L3 cache to 4MB and the system cache to 6MB, boosting performance for large applications.

The older Snapdragon 850 features a dual-core Adreno 630 GPU, but the 8cx shifts to a quad-core design based on the same microarchitecture. The remaining 8cx features are similar to those of the 855. They include the same Hexagon 690 DSP, with its new tensor engine for AI acceleration; the same image processor (ISP), audio, and video engines; the same Wi-Fi and Bluetooth connectivity; and the same LTE modem, rated at an industry-leading 2.0Gbps.

The company expects the first 8cx PCs to ship in 3Q19. In the meantime, it'll continue to offer the 850, which recently began shipping in a pair of 2-in-1 Windows devices: the Lenovo Yoga C630 WOS ($699) and the Samsung Galaxy Book 2 ($999). These systems provide longer battery life than Intel-based 2-in-1s and include an LTE modem for always-connected capability, should the customer choose to pay for monthly service. The downside, according to PC reviewers, is limited performance, particularly for apps that haven't been ported to Arm.

Microprocessor Report subscribers can access the full article:

Barefoot Joins 400GbE-Switch Club

By Bob Wheeler

Barefoot Networks plans to bring programmability to 400G Ethernet (400GbE), and its past performance suggests it can. This month, it preannounced its Tofino 2 second-generation switch chip for 1H19 sampling. Employing 50Gbps PAM4 serdes, the new product delivers 12.8Tbps of bandwidth -- double that of the first Tofino. Like its predecessor, Tofino 2 is programmable using the P4 language. It's also the first announced 12.8Tbps switch using 7nm technology, promising lower power than 16nm designs achieve.

Tofino 2 uses the same number of pipelines as the shipping Tofino. The only architectural change between generations is packet-processing extensions (dubbed PPX) to the match+action units (MAUs), but Barefoot withheld details. The company created three Tofino 2 families: the U (ultra) family provides the most MAUs and associated memory, the M (mainstream) family implements fewer MAUs, and the H (hyperscale) family omits PPX and offers the lowest latency. The 12.8Tbps versions handle 32x400GbE, 128x100GbE, and 256x50GbE ports. Barefoot will also sell 8.0Tbps and 6.4Tbps versions.

Tofino 2 samples will trail those of Broadcom's 12.8Tbps Tomahawk 3 by about 18 months. Furthermore, Broadcom and Innovium have both started production of their respective 12.8Tbps chips, enabling the first high-density 400GbE switch systems. On the plus side, Barefoot's use of 7nm technology should deliver superior power dissipation for the new design. The startup has demonstrated that a programmable switch chip can directly compete with a fixed-function device, minimizing any programmability "tax."

Microprocessor Report subscribers can access the full article:

Year in Review: IP Suppliers Push AI Accelerators

By Mike Demler

In 2018, to meet increasing demand for on-device AI such as face and voice recognition, intellectual-property (IP) vendors added neural-network functions to all types of processor IP, from CPUs, DSPs, and GPUs to dedicated deep-learning accelerators (DLAs). Several vendors rolled out DLAs with large multiply-accumulate units, which boost neural-network performance, and some introduced multicore configurations. These licensable cores enable AI to expand from data centers to the edge, serving in IoT sensors, smart speakers, smartphones, and many other client devices.

In traditional CPU IP, the open-source movement that RISC-V ISA proponents began shows no signs of slowing. Designers can choose from a growing list of vendors offering verified cores. First-generation RISC-V CPUs targeted embedded microcontrollers, but new generations compete with some of Arm's Cortex-A CPUs. The open-source business model is catching on with other companies as well. Wave Computing announced plans to open the MIPS Release-6 ISA, and Nvidia opened the DLA in the Xavier autonomous-driving processor.

Although RISC-V is an alternative to some of Arm's low-performance tiers, Cortex-A CPUs are moving into market segments that were previously the exclusive domain of x86. The company's new Cortex-A76 powers Windows-on-Arm PCs, and Amazon announced plans to use custom Arm-v8-based processors in its cloud-computing service.

In 2019, we expect most DLA vendors will include weight compression and variable precision as standard features, shifting their emphasis to greater area and power efficiency. In the coming year, Arm will embark on its new strategy, adding to its traditional Cortex portfolio but also launching two other product lines: "automotive enhanced" cores and the Neoverse family, which is optimized for data centers and infrastructure.

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. To subscribe, please visit:

Domain: Electronics
Category: Semiconductors

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