ARM Preps Near-Threshold Processor for IoT

http://www.eetimes.com/document.asp?doc_id=1319229&elq=4a6f945594cc4d8a87cb22309804618a

 

ARM Preps Near-Threshold Processor for IoT

LONDON — ARM Holdings plc is working on a processor core

optimized for operation close to the threshold voltage of CMOS transistors

and at clock frequencies of the order of tens of kilohertz.

The processor IP licensor has done multiple internal designs in research

and is now developing a possible slow,

very low-power microcontroller core for the Internet of Things (IoT) and other embedded applications.

Mike Muller, chief technology officer of ARM,

told a group of UK journalists that strategies for processing small amounts of data

and transmitting small payloads for the IoT might differ depending on whether a node was powered

or relied on energy scavenging from the local environment.

 

For the latter case, a processor optimized for exceedingly low-power consumption

that could wake up whenever enough energy had been harvested could be useful.

"Normally,

the best strategy is to do processing as fast as possible and then go to sleep for as long as possible

— get in and get out," Muller said.

"But for energy scavenging, it can be different."

In these cases, it may be best to power up and get the packet out at minimum energy as soon as enough energy has been harvested.

"And it turns out the design tradeoffs are different."

Mike Muller, CTO of ARM.

Mike Muller, CTO of ARM.

His Cambridge, England, company already has a reputation

for low-power processor designs in mobile phones, tablets,

and embedded applications.

But the IoT — which may include larges numbers of wireless sensor nodes communicating

with one another and host networks

— may require further improvements in the power-performance ratio.

 

One way to achieve that is to take the voltage down close to

or even below the threshold where a transistor can be turned on and significant current flows

— in the region of 0.3-0.6V.

 

Though transistors can be characterized at these near- and subthreshold voltages

and consume much less energy there, switching times are longer, making performance slower.

Three years ago, when EE Times asked Muller about the possibility of subthreshold design,

he expressed scepticism.

If you really want to save power, the laws of physics may say it is a good idea,

but you are still trading off performance against voltage;

you could only get a few hundred kilohertz of clock frequency.

And because there is no major demand, it is not possible to get foundries to produce qualified silicon.

 

However, Muller now says ARM has been working on the technology since then.

We did a couple of subthreshold designs and then a bunch of near-threshold designs.

Near-threshold is easier, because you can be closer to the SPICE models.

We did some library development to go with that.
It turns out there are some real implementation choices to be made when you design slow processors;

where and how do you do power gating, where do you do clock gating.

We’ve come up with some techniques for power gating in the [clock] cycle.

And then there are things like drowsy logic.

Drowsy logic, an alternative to putting the processor into sleep mode,

is applicable where leakage currents are managed.

Muller said that it is possible to turn off the power and let the power rail voltage decay.

But it is not necessary to provide explicit state retention as long as the voltages

on flipflops do not go too low.

Essentially, it is a method of intermittently providing voltage sufficient to maintain state and operations.

There are multiple techniques, including clock gating,

and how these techniques are best applied depends on the clock frequency being targeted.

"With the clock gating 20kHz is very different to 2MHz or 2GHz."

The advantage of selecting a near-threshold voltage is scalability from conventional CMOS operations,

making it more likely that a foundry could provide support,

Muller said.

Companies have designed circuits to operate at subthreshold voltages,

but they have to take all the risk in terms of characterizing and testing their silicon.

"There are very few companies who are prepared to do that. Near-threshold is more compatible with the foundries."

When asked if ARM was working with particular foundries to offer near-threshold manufacturing processes,

Muller said:

"That’s all part of market development."

The technology is compatible with the Cortex-M0, an ARM 32-bit core used in low-power microcontrollers,

Muller said, adding:

"It’s all about the implementation"

When asked about the status of the research,

Muller said it had moved into development with partner companies

and would then go to implementation.

He did not provide a timetable.

————————————————————————————————————————————————–

發佈留言

發佈留言必須填寫的電子郵件地址不會公開。