WP_Term Object
(
    [term_id] => 13
    [name] => Arm
    [slug] => arm
    [term_group] => 0
    [term_taxonomy_id] => 13
    [taxonomy] => category
    [description] => 
    [parent] => 178
    [count] => 389
    [filter] => raw
    [cat_ID] => 13
    [category_count] => 389
    [category_description] => 
    [cat_name] => Arm
    [category_nicename] => arm
    [category_parent] => 178
)
            
Mobile Unleashed Banner SemiWiki
WP_Term Object
(
    [term_id] => 13
    [name] => Arm
    [slug] => arm
    [term_group] => 0
    [term_taxonomy_id] => 13
    [taxonomy] => category
    [description] => 
    [parent] => 178
    [count] => 389
    [filter] => raw
    [cat_ID] => 13
    [category_count] => 389
    [category_description] => 
    [cat_name] => Arm
    [category_nicename] => arm
    [category_parent] => 178
)

Processors, Processors, Processors Everywhere

Processors, Processors, Processors Everywhere
by Tom Simon on 10-06-2016 at 7:00 am

At first glance a processor conference might seem a bit arcane, however we live in an era where processors are ubiquitous. There is hardly any aspect of our lives that they do not touch in some way. Last week at the Linley Processor Conference the topics included deep learning, autonomous driving, energy, manufacturing, smart cities, commerce and more. The conference was led off by a keynote from the conference’s namesake Linley Gwennap, who touched on all the main themes for the following two days.

The keynote presented many familiar topics and ideas, along with several surprising ones. Let me summarize the most interesting points.

Linley observed that increasing wafer costs have reached the point where the price per transistor is actually going up. 20nm was the crossover point for this. As Linley puts it – “Moore’s Law is only for the rich.” The effect of this will be that cost sensitive products will stay at 28nm. Thus mainstream products will be limited in the amount of integration they contain. However, high end products will continue to move to new advanced nodes because the justification for higher prices exists.

During the era of declining transistor costs, processors were in a race that demanded a new release every two years, which starved necessary architecture changes. We saw a continuous stream of general purpose processors as a result. This is likely to change. Specialized architectures can offer a 10 or even 100 times improvement in performance-per-watt metrics. We see numerous examples of this from companies such as Tensilica, or Microsoft with their “Catapult” project which combines an FPGA with a general-purpose processor. The other trend that this will drive is the addition of more specialized processing in accelerators to offload CPU’s.

Vision processing has become a very active area because it has widespread applications. Vision processing is being used in gaming, mobile devices, industrial applications and advanced automotive. Vision Processing Units (VPU’s) which serve this market are available from Cadence, Ceva, Synopsys, and VeriSilicon. Large verticals such as NXP and Intel are entering this market through acquisitions.

Neural networks are also changing the processor landscape. Neural network training requires massive data bandwidth and high precision floating point processing. Whereas, the recognition process needs highly parallel but smaller data size processing. Many players are active in this market. Google has developed a special purpose ASIC for TensorFlow. There is also Wave Computing, Intel with its acquisition of Nervana Systems, as well as IBM and others.

Linley pointed out that data center growth has been very good for Intel. This market has grown by 11%. Interestingly the public cloud which includes Amazon Google and Alibaba is the fastest growing segment. Intel’s Xeon E5 has become the mainstream processor for 2S systems. This is now using the 14nm Broadwell-EP. Also there is the Xeon D which offers a single socket SOC with up to 16 Broadwell cores. Avoton initially targeted low-cost micro servers with its eight Atom cores. However, the interest in micro servers has diminished leaving Avoton to target the embedded market.

At the same time Intel is seeing challengers for Xeon. IBM has its Open Power initiative, which is bearing fruit with the several Power8 processors already available in servers. And in 2017 we can expect to see Power9 based servers. Also AMD is moving forward with a new Zen CPU.

There is also a bevy of activity in ARM based alternatives to the x86 architecture. AppliedMicro is seeking advantage with X-Gene 1 and 2. Cavium is focusing on high throughput with their upcoming ThunderX. QUALCOMM and Broadcom could also deploy new ARM server processors in 2017.

ARMv8 processors are moving into the embedded space. For instance, AppliedMicro and Cavium both offer embedded versions of their multicore ARMv8 server processors. QUALCOMM is likely to do the same on designs using over 24 ARMv8 CPU cores. NXP is now sampling QorIQ LS2 with up to eight Cortex-A72 cores. And the Broadcom StrataGX line moves to Cortex A57.

The usual distinction between server processors and embedded SOC’s is beginning to blur. Some Intel processors are suitable for both server and networking designs. Activities such as network function virtualization are starting to use server infrastructure. What were previously network processors are starting to look like multicore embedded processors where we see SMP Linux and GNU tools. This means that some of these network processors come with vastly improved development environments. The winning future network architecture is in flux.

One of the big stories from this conference was how network function virtualization is moving from the core to the edge. This is enabled by new development tools and changing hardware. We are seeing intelligent network adapters that include processors or FPGA’s that offer a wide range of programmability. This makes it possible to offload a large number of tasks. In some cases, virtual switch functions are running in virtualized servers on the NIC. This will do a lot to improve performance inefficiency.

Linley feels pretty strongly that the real market for IOT is business. There is high motivation in business for the improvements that IOT can bring. Anywhere there is cost savings there is a reason to innovate. IOT offers a compelling business case because of the many ways that it can improve process efficiency, conserve resources, improve security and grow markets. We can already see it being used in smart meters, parking, lighting, energy, vending machines and more.

The consumer and home market will certainly grow as well, especially as this technology gains momentum. Consumers are motivated by convenience, time savings, security, as well as status and social engagement. However, the direct economic benefits it brings are smaller and less tangible.

Processors will play a central role in making IoT data secure, which is an essential prerequisite for market growth. Linley emphasized that IoT data needs to be secure during transmission. With the ease of interception of wireless data, it needs to be encrypted. Stored data at rest in the cloud is also potentially vulnerable. Cloud service providers need to take precautions. IoT sensor and edge devices are vulnerable to hacking because they are difficult to physically secure. This is where secure boot comes into play. Also secure on-chip storage for crypto keys is necessary. Lastly, there needs to be a way to authenticate incoming commands.

Next he spoke about how cars are now built containing dozens of processors. Microcontrollers are used for things like windows, wipers etc. Then there are the processors used for in-dash electronics. Applications include navigation, user interface as well as digital dashboards and the surround view video. Pile on top of this the processing needs for ADAS, and we can easily see why automotive is a huge and growing market for processors. Linley sees this market at $10 billion annually and growing. In fact, he is suggesting that it could double by 2025.

For a moment let’s look at the requirements for ADAS. The players in this market are names from the smartphone processor market. Nvidia, NXP and TI are all recognizable with Tegra, i.MX and OMAP, respectively. Though, most of these need a vision processing engine to boost performance in the ADAS application. This is where Neural Networks come into play. For recognition massively parallel 8-bit operations are needed. For ADAS there is also a sensor fusion processing requirement of the highest order. Radar, Lidar, ultrasonic, infrared and optical all need to be combined to create the internal 3D virtual world the ADAS system will use to make effective and safe operational decisions.

Already Tesla, Volvo and others offer driver assist, which in many cases does an impressive job. These systems require driver supervision, but greatly aid in reducing driver workload. The first wave of autonomous vehicles could hit the market in 2018. At the very least they will be built with the processing power, but may lack the final software. Ford recently announced that it plans to produce a fully autonomous vehicle with no steering wheel in 2021.

While these systems may add $5,000 or more to the BOM for a car, businesses like Uber or Lyft would stand to see huge net savings if they can eliminate the cost of drivers. I know this all seems like science fiction, but we are poised on a dramatic precipice. My own thought on the rapid progress in autonomous vehicles is that it stems from the exceptionally heavy traffic found in Silicon Valley. Nowhere else will you find the motivation and the talent together to accomplish such a difficult task.

After Linley’s keynote there were two days of detailed presentations on a wide range of topics, including those above. For more information about this and other Linley conferences follow this link.

Share this post via:

Comments

0 Replies to “Processors, Processors, Processors Everywhere”

You must register or log in to view/post comments.