Startup tips application-specific video processor

Startup tips application-specific video processor

EETimes

Startup tips application-specific video processor
By By Peter Clarke, EE Times
May 23, 2001 (4:33 p.m. EST)
URL: http://www.eetimes.com/story/OEG20010523S0087

LONDON — The VideoCore chip, a 16-engine application-specific processor able to deliver audio and video data stream processing, including MPEG-4, is in the driver's seat at startup Alphamosaic Ltd. The chip is intended for use in portable equipment and mobile phones that could offer security video of remote sites, send video postcards or support mobile videoconferencing.

Alphamosaic, which started a few weeks ago as a spin-off from Cambridge Consultants Ltd. (Cambridge, England), said the device offers the flexibility of a programmable solution yet rivals the power consumption of a hardwired, standard-specific engine such as an MPEG-4 decoder. "We've taken a RISC philosophy and crossed it with a parallel-processing approach," said Robert Swann, co-founder and marketing director of Alphamosaic.

Each of the VideoCore's 16 engines "is a fully operational 32-bit processor, but each is optimized to do the kind of math that lies behind all video -processing algorithms — the sort of thing that occurs again and again within the DCT [discrete cosine transform]," Swann said. "The secret to computational efficiency is the way you control the engines. It's very important that you keep them busy all the time."

GSM videophone

Though Alphamosaic is a brand-new company, the processing engine at the heart of the VideoCore chip was designed when most of its 12 employees comprised the video-processing design team within Cambridge Consulting.

The team's most significant project to date was developing a Global System for Mobile Communications videophone that can send and receive color quarter-CIF pictures over a 28.8-kbit/second, high-speed circuit-switched data GSM network. That work was done for Orange, a United Kingdom service provider, and the product launched at the end of 2000.

The team is now completing the first VideoCore video-application-specific processor implementation. "The core [of VideoCore] has been designed, but we are aiming to have a product for external use by the beginning of next year," said Swann.

He added that the next few months will be spent deciding on the peripherals, ports and memory caches that will be needed, and implementing the complete chip for manufacture at a Southeast Asian foundry. Swann said Alphamosaic is targeting a 0.13-micron CMOS process technology able to support embedded memory.

Alphamosaic included two layers of software with its hardware design, Swann said. The first level is a set of software primitive functions that are frequently used within video processing. These are written at an "assembler macro" level, he said, to make the best use of the multiple engines available. The second level is an ANSI C compiler that can take the C-language code and compile it to call the assembler macros and drive the 16-engine VideoCore directly.

Each of the engines and the VideoCore as a unit can handle data word sizes up to 32 bits.

Swann said that 16 is a near-optimum number of proce ssing elements. More than that and the problem becomes one of interengine communications; less and you don't really get much benefit of scaling, he said.

"We've already got an MPEG-4 codec running, but we can also run other people's codecs, and the programmability gives users flexibility in case a standard should change or a new standard should emerge," said Swann.

According to simulation at a given clock frequency, he said, the VideoCore was 40 times more efficient at DCT than a StrongARM RISC and 160 times more efficient at performing certain types of mathematic morphology operations that are useful for identifying objects in scenes. At the same time, the core uses considerably less power — typically, less than 100 milliwatts — than a standard RISC, providing greater power efficiency for video processing.

Copyright © 2003 CMP Media, LLC | Privacy Statement
×
Semiconductor IP