Samsung, Intrinsity pump ARM to GHz rate

July 24, 2009 //

Samsung Electronics Co. and Intrinsity Inc. announced they have first silicon for Humming bird, an ARM Cortex A8 that runs at a GHz and delivers more than 2,000 Dhrystone Mips while consuming 640 mW power, competing with Intel Atom processors in an expanding market of mobile devices.

SAN JOSE, Calif. — Samsung Electronics Co. and Intrinsity Inc. announced they have first silicon for an ARM Cortex A8 that runs at a GHz and delivers more than 2,000 Dhrystone Mips. The Hummingbird chip arrives at a time when ARM and Intel Atom processors are in increasing competition for an expanding market of mobile devices.

Currently ARM dominates the market for smart phones such as the Apple iPhone. Intel is gearing up its x86-based Atom chip to address that market although in its current form it still consumes as much as 2W at 2 GHz, too much for cellphone handsets.

"The new personal computer is the thing you carry in your pocket that has communications capabilities and can browse the Internet," said Tom Halfhill, senior editor of the Microprocessor Report. "Both ARM and Intel know that, and they want to be in those devices," he said.

While average selling prices for processor are much lower in cellphones than PCs, volumes are significantly higher. Will Strauss of market watcher Forward Concepts (Tempe, Ariz.) estimates as many as 1.2 billion cellphones will be sold next year compared to about 300 million PCs.

Hummingbird is a natural follow on for today's Apple iPhone 3GS and the Palm Pre handsets using ARM Cortex A8 processors running at up to 600 MHz from Samsung and Texas Instruments respectively. Samsung has not commented on its plans for the new core.

Samsung and Intrinsity have booted operating systems and run customer applications on first silicon of Hummingbird which could be in production before the end of the year. Intrinsity provided custom circuit technology that powers the chip, built in a low power 45nm process.

The companies estimate Hummingbird consumes less than 0.75 mW/MHz and power leakage is "in the single digits." Intrinsity said it has most of the technology needed to help other ARM licensees build similar processors.

"We have customers looking to put this part in a TSMC 40nm process, and we could have a part available in less than six months," said Bob Russo, chief executive of Intrinsity. "We'll probably have the fastest [A8] part manufactured to date with the best megahertz per Watt and the lowest leakage," he claimed.

Intrinsity also is at work applying its proprietary circuit technology to other mobile processor designs across a variety of architectures. The projects generally target data rates of a GHz or beyond, and one design uses two cores.

"I don't see any end to the frequency race" for cellphones, said Russo.

Neither does Intel. The PC giant said it is designing a 45nm Atom called Lincroft aiming at next-generation mobile devices including smart phones. In 2011 it plans to release Medfield, a 32nm system-on-chip based around an Atom core.

"We are very excited about our 32nm generation of SoCs," said Gadi Singer who manages Intel's mobile processor operations and its SoC enablement group.

Comparing ARM Cortex A8 and Atom processors is no easy job.

"Cortex A8 is the highest performance ARM core, and Atom is the lowest power consumption x86, but they are still pretty far apart," said analyst Halfhill. "ARM is moving into the direction of higher performance while Intel is moving into lower power, and the current collision point seems to be netbooks and mobile Internet devices," he said

"The initial Atom is still probably four times bigger in die area and consumes three times more power than a Cortex A8," said Linley Gwennap, principal of market watcher the Linley Group (Mountain View, Calif.). "Atom probably is two or three times the performance of Cortex depending on what clock speed you are using, but it doesn't matter how fast it is if it doesn't fall into power budget of a smart phone," he added.

The two processors share a similar superscalar architecture although ARM's is 32-bit wide compared to a 64-bit width for Atom. The two have radically different instruction sets, although both support DSP-like extensions for media processing.

"Overall, I don't think architecturally there's a huge difference" between ARM and Atom, Halfhill said.

Gwennap predicted Intel will need both 32nm technology and major architectural changes to get Atom into the power budget of a smart phone. "That's another one or two generations down the road," he said.

"We know [Intel] can get to a GHz and beyond, but typically they quote their best power at 500 MHz, so it's hard to do an apples-to-apples comparison," said Dave Shippey, vice president of engineering at Intrinsity. "We think our power is much lower, and I am sure our leakage is lower because there is less silicon in an ARM implementation, but I don't have a direct comparison," he added.

An Intel spokeswoman admitted the company has yet to release specific numbers on Atom performance and power consumption. However, the company has said its next-generation Moorestown platform slated for a 2010 launch will have a 50-fold improvement in system-level idle power while cutting board size in half.

The 2011 Intel chip, Medfield, will sport further reductions in power and size "and will extend Intel solidly into smart phone segments," she added.

TI estimates Intel's Moorestown platform which consists of the Lincroft processor and a Langwell peripheral chip will consume an estimated 1.5W and take up 392 square millimeters of packaged chip space. By comparison, TI's single-chip OMAP3640 will consume 0.5W and take up 144 square millimeters, said Brian Carlson, a TI marketing manager.

In addition, TI will sample this fall an OMAP4 chip that puts two ARM Cortex A9 cores in the space of a single Atom core, he added.

"I've looked out to the 22nm node on both sides, and I still see a gap--the gap narrows, but I wouldn't say we are at parity," said Carlson. "The x86 will always have more transistors, and if they lose [transistors] they will lose software compatibility or performance," he said.

The Intel spokeswoman said the TI comparison was flawed on several fronts. It does not take into account video or graphics performance, and TI should compare current Atom Z-series parts to TI's current OMAP 3440, she said.

Perhaps the biggest difference between ARM and Atom lies in the business models of the two companies.

ARM licenses its designs to all comers as synthesizable soft cores. Some big customers can negotiate architectural licenses—rumored to cost as much as $20 million--to define their own derivative cores.

By contrast, most Atom chips will be made by Intel itself. "Intel's worry has to be there are five to ten companies with ARM-based products coming after the space Atom occupies--and they will be hitting the streets this fall," said Strauss of Forward Concepts.

"The ARM licensing model is a force multiplier," said Halfhill. "All the Atom innovation has to come from Intel," he said.

Taking a small step toward the ARM model, Intel struck a deal with TSMC in March so third parties can make SoCs using an Atom-based core Intel is now porting to a TSMC process. To date the companies have divulged few details about the effort.

Unlike the RTL-based ARM cores, the Atom TSMC core is expected to be defined in hardware. That leaves SoC designers far less flexibility, although they will potentially get an Atom core closer in performance to Intel's own designs.

"It's a more restrictive model at best," said Halfhill. "My understanding is Intel controls the [TSMC-based SoC] designs," he added.

Intel is itself coming up to speed as an SoC provider. It shipped its first chips about a year ago and has 14 SoCs in the works for its 32nm process.

"There's a huge [SoC] learning curve, and they are going down it but there is still significant advantages going down the ARM side," said Carlson of TI.

Intel's 14 32nm SoC designs is "a big deal, but only Intel gets to use that process," said Strauss.

Meanwhile the competition among ARM suppliers is significant and growing.

Halfhill said Freescale, Marvell, Samsung and Qualcomm have ARM architectural licenses. Apple probably has an architectural license, too, he speculated, based on the company's efforts to build its own silicon design team.

Qualcomm announced in November a dual-core version of its Snapdragon chip for mobile devices. ARM and another company are also working on a multi-gigahertz, multi-core ARM processor that is probably a year away, one source said.

"Marvell has been claiming they can crank up its Sheeva ARM-based processor to 2 GHz on a good day with a tail wind, so it's possible Marvell could put together a multicore muti-gigahertz ARM processor," said analyst Gwennap. "That would be pretty competitive with what Mips Technologies is doing and what Intel may be working on," he said.

Russo of Intrinsity said Marvell spent more than $100 million developing Sheeva and Qualcomm spent nearly than four years and $300 million on its Snapdragon design.

"We did [Hummingbird] in 12 months for a small fraction of that cost," Russo said. "We can turn these cores fairly quickly with a staff of 40-70 people on it for most of the time," he said.

Technical papers     

• High Voltage Surge Stoppers Ensure Reliable Operation During Power Surges
• AR RF/Microwave Instrumentation Field Analyzers
• 16-Bit, 100 kSPS Low Power Successive Approximation ADC System
• Replacing the Candle in the Candelabra

December 15, 2011 | Texas instruments | 222901974

Unique Ser/Des technology supports encrypted video and audio content with full duplex bi-directional control channel over a single wire interface.