Hot Chips Goes Deep (Learning)

Hot Chips, where the semiconductor industry struts new stuff, focuses on the machine learning and the maker movement, with keynotes from Facebook and Qualcomm and announcements expected from some interesting smaller companies.

Every August I look forward to the Hot Chips. It’s the conference I can rely on to bring together most of the interesting developments in microprocessor development all under one roof. The Hot Chips program committee does an exemplary job of keeping pace with industry trends and academia research.

This year the committee emphasizes machine learning with a Sunday morning tutorial, the Monday keynote, and a presentation from Microsoft Research. The Sunday tutorial will cover an overview of the architectures, algorithms, and computations used in commonly deployed machine learning models and will describe current challenges and research directions. On Monday, the keynote address from Facebook on its research into machine learning with convolutional neural networks, to be presented by Yann LeCun.

A second tutorial on Sunday will address the rise of the maker movement. Of particular interest is section from Adapteva, which boot strapped its unique compute architecture chip in part using Kickstarter. ARM will cap that session with an overview of trends for hardware and software developers.

The Tuesday keynote brings in Qualcomm’s CTO Matt Grob to talk “The Road to 5G.” The need for more wireless throughput is unrelenting and the keynote presentation will look at the mega trends driving that need for more bandwidth and the technical hurdles needed to be overcome. I expect we’ll hear about the complex signal processing needs required for spatial reuse (beamforming, MU MIMO, etc.) of limited spectrum.

Last year’s event had a tutorial covering IoT and this year there’s a session that features presentations that represent three vectors of IoT: a fully integrated platform from ST Micro, Intel’s IoT development for wearables, and an ultra-low-power chip from PsiKick that can run solely on energy harvesting (no battery).

Hot Chips often features specialized processors for media processing and DSP. Many heterogeneous SoC processors include DSPs and multimedia processors for image and signal processing. Because of the Qualcomm’s standing as a leader in smartphone designs, its DSP architecture is pervasive and the company will talk about the latest V6x Hexagon DSP used for mobile imaging and always-on applications. TI will talk about its entry into the hot automotive advanced driver assistance system (ADAS) using a heterogeneous, multicore architecture. AMD’s “Carrizo” PC processor has extensive graphics and multi-media capabilities for notebook and desktop computers. A non-traditional approach to DSP design using cores will be discussed by Kalray.

For cloud and high performance computing there’s updates from Applied Micro on its X-Gene roadmap and Oracle will be talking about a new SPARC processor called Sonoma. The most intriguing presentation in this session is the “Mars” processor from unknown Chinese vendor Phytium that will use 64 ARMv8 cores. IBM will provide a view into the accelerator architecture built into the Power8 processor which allows FPGA and GPU attachments.

One very competitive silicon market is FPGAs. At Hot Chips, both Altera and Xilinx will be talking about their FinFET generation designs where the saving in power per transistor has been used to increase speeds — now approaching 1GHz. In addition, the extra transistors can be used to add helper cores to the programmable logic. This also the session where Microsoft Research will describe how they’ve using FPGA logic to accelerate deep learning and how to scale out those designs. The presentation will provide insight into how FPGAs can be used for dynamic learning algorithms where previous Microsoft presentations on FPGAs focused on fixed function acceleration.

The session on Graphics processing units (GPU) features AMD’s newest high bandwidth memory design presented by AMD’s CTO Joe Macre. The Radeon R9 Fury graphics card is the first volume product using high bandwidth memory which combines stacked memory and a silicon interposer. ARM will present details of its new high-end Mali-T880 Mobile GPU IP. Academia, in this case University of Wisconsin, takes a shot at building a GPU design, but not for graphics. The MIAOW project is an open source design that can be used for GPGPU (GPU compute) application using the OpenCL API.

There’s always some presentations that don’t fit neatly into neat categories. This year in the “Applications” sessions is a potpourri of three interesting technical topics: a chip designed for professional 4K and 8K broadcast video using the latest H.265/HEVC video codecs; a CMOS biosensor that has ultra-low-light capabilities to track diseases; and a multispeed Ethernet PHY that supports 2.5Gbps and 5Gbps rates over legacy copper cables.

The last session on Tuesday is traditionally the main “big” processor session. This year the session features three Intel processors ranging from the 2nd generation of the many-core Xeon Phi Processor for HPC, the scale-out server Xeon D processor, and the low-power Intel “Cherry Trail” Atom processor. While none of these are new processors reveals, it’s expected that Intel engineers will provide more details of each, including design tradeoffs. The odd duck in the session is an implementation of UC Berkeley RISC-V Vector Processor. Last year the Berkeley contingent showed off RISC-V instruction set in the break area, but now with a real chip, they made it to inside the auditorium. It’s not too often you see a chip design of this integration and complexity coming from academia. What started as a project to give universities a royalty-free and extendable CPU architecture to build on, has gained traction, especially in India and Asia for development purposes.

Once again Hot Chips offers a compelling mix of academically interesting projects with industry leading commercial designs. Each year the event draws the top talent in microprocessor design with compelling content in a collegial atmosphere. For more information on registration and location, please check out the website at www.hotchips.org. Come for the technical content, stay for the conversations.

—Kevin Krewell is principal analyst at TIRIAS Research.




Source: Hot Chips Goes Deep (Learning)

Via: Google Alert for Deep Learning