Computex has always been a PC show and with good reason--Taiwan designs most of the world's computers and manufactures many of the components that end up in them. But with the PC market still in the doldrums, this year there is a concerted effort to broaden the show to encompass buzzier themes such as the Internet of Things, artificial intelligence and robotics, and virtual reality.
That trend was reflected in Intel's opening keynote, which covered a wide range of technology from home gateways to drones to VR gaming to media processing to machine intelligence. That's not say the PC wasn't present. But Intel rushed though the PCs mentioning a handful of innovative PCs (HP Spectre 13, Lenovo Ideacentre 610s and Intel Skull Canyon NUC) and 2-in-1s (Acer Switch 12, Dell Inspiron 11 3000 and Huawei MateBook). Asus Chairman Jonney Shih made it on stage--as he seems to every year--but had to plead for more time to mention a few features of the Transformer 3 and Transformer 3 Pro. It is telling that Navin Shenoy, the head of the Client Computing Group, felt the need to reassure the audience that Intel is "resolute and unwavering in our support for the PC."
Intel did announce a new processor, but it is a niche product. The Broadwell-E Core i7 Extreme Edition is the company's first desktop chip with up to 10 cores (20 threads), along with 25MB of last-level cache, four memory channels, 40 lanes of PCIe and Thunderbolt 3.0. It also includes a new version of Turbo Boost that Intel says can squeeze more performance out of individual cores (the 10-core Core i7-6950X has a base frequency of 3.0GHz but can burst up to 3.5GHz). Intel is positioning the chip for "mega-taskers"--a category that combines 4K gaming with other intensive background tasks--but the vast majority of users will get better value from high-end quad-core Skylake chips such as the 4.0GHz Core i7-6700K.
The more mainstream Apollo Lake for entry-level PCs, 2-in-1s and tablets as well Kaby Lake, Intel's third 14nm product family with some 400 designs in the wings, are both in production and will be available by the end of the year. Shenoy also confirmed that Intel's Optane solid-state drives based on 3D XPoint memory will be in production by the end of the year and showed how they can deliver better rendering performance than SSDs using conventional NAND flash memory.
But much of Intel's focus at Computex is on other areas, part of a strategy that CEO Brian Krzanich recently described in the wake of announcing a restructuring as "transforming Intel from a PC company to a company that powers the cloud and billions of smart, connected computing devices." During the keynote, Intel announced an Atom-based SoC and Wi-Fi chipset designed to support the growing number of connected devices in the home all demanding more bandwidth. Shenoy predicted that IoT will transform not only the home, but also transportation, manufacturing, surveillance and healthcare. The company talked about a pilot project with Taiwan's SanJet Technology and Chunghwa Telecom to collect real-time data on drivers in Taipei and transmit it via 3G wireless to the cloud for analysis (think of those Progressive Snapshot ads).
One device Intel really loves is the drone. They have a prominent role at nearly every recent Intel event and Krzanich heads up the FAA's drone advisory council. Most of the attention has been on consumer drones such as Yuneec's Typhoon H, which uses Intel's RealSense 3D cameras to automatically detect and avoid obstacles. But Shenoy said Intel also sees a large opportunity for commercial drones to handle tasks such as surveying construction sites or inspecting power lines. At the recent Intel Developer Forum in Shenzhen, Intel announced the Aero platform for developing UAVs (unmanned aerial vehicles) with advanced features.
Intel believes that connected devices will drive demand for more networking and datacenter infrastructure, which will in turn support more devices in a virtuous cycle. "Today's networks are certainly not capable of supporting the 50 billion devices to come," Diane M. Bryant, Executive Vice President and head of the Data Center Group said at the Computex keynote. She added that networks need to become virtualized, software-defined, and cloud enabled--a trend the industry refers to as NFV or network functions virtualization. Foxconn's Fang-ming Lu joined her on stage to announce a partnership to develop NFV gear.
Much of the increased network traffic is video. Intel said it currently accounts for 72 percent of all Internet traffic growing to more than 80 percent over the next few years. The company announced the Xeon E3-1500 v5 Skylake with Iris Pro P580 graphics, which it says is a more cost-effective solution to ASICs, DSPs and programmable logic for cloud and communication service providers that need to deliver video in multiple formats. The E3-1585L v5 with hardware-assisted HEVC encoding can deliver two real-time 4K 30fps HEVC streams or eight 1080p 30fps HEVC stream, and it can simultaneously transcode 18 AVC 1080p 30fps streams to HEVC at the same resolution. HEVC is a compression standard that can deliver high-quality video at lower bit rates, but it requires significantly more processing power.
In addition to more video, Bryant predicted that the kinds of video will change. The so-called Visual Cloud 2.0 will include immersive virtual reality and augmented reality. To illustrate this, start-up Rivet demonstrated its 360-degree video service, which uses HP Moonshot servers with Intel Xeon E3 processors, with a live performance of the band Living Colour from the Blue Note Jazz Club in New York. It didn't quite go as planned--the band happened to be on a break at the wrong time--but the concept of remotely experiencing a concert or other event in an immersive way seems promising.
With the recent advances in artificial intelligence, starts-ups and established semiconductor companies such as Nvidia and IBM have been rushing to develop new chips for deep-learning applications. Intel has been relatively quiet, however, so it was interesting to see Bryant talk about using Xeon for machine learning. Don Hsi, the chairman of an interesting start-up called Viscovery, explained how the service used QCT's QuantaPlex servers with Knights Landing Xeon Phi processors to recognize and tag objects in videos for search or e-commerce applications. Bryant said Knights Landing delivered 40 percent better performance than GPUs for this visual search application.
Intel is pushing the idea of a single architecture and programming model for machine learning using Xeon Phi for training the models and either Xeon, for best performance per dollar, or Xeon and Altera FPGAs, for the best performance per watt, for "scoring" or inferencing, terms for running the models. But they face stiff competition here from Nvidia, which has a big lead in training with most major cloud service providers already using Tesla GPUs, and is making progress on inferencing with the Tesla M4 and Drive PX2 module. There's plenty of other competition for inferencing too. Google has designed its own Tensor Processing Units, and Cadence, Freescale, Mobileye, NXP, Synopsys, and start-ups such as Movidius and Nervana are all vying for a piece of this growing market.