The Chromatic Research Mpact Media Engine

A Completely New Approach to Multimedia Integration

Chromatic Research is redefining the personal computer with a completely new approach to delivering multimedia capabilities that promises for the first time to bring a full range of high-performance multimedia functions to the mainstream. For the same or lower cost as today's most basic multimedia functionality, the Chromatic Research Mpact media engine will enable PC manufacturers to deliver digital video, superVGA 2D graphics acceleration, 3D graphics acceleration, CD-quality stereo audio and realistic sound effects, fax/modem capabilities, a range of telephony options and desktop videoconferencing. Using a single system design, PC manufacturers can provide any or all of these functions through simple software upgrades that can be implemented either before or after the system is purchased.

The key to this breakthrough is Chromatic's Mpact media engine, a revolutionary departure from the traditional way of providing multimedia capabilities. Rather than integrating existing fixed-function, board-level solutions onto a single chip -- which leads to higher transistor counts, poor transistor utilization and larger off-chip memories -- Chromatic Research has designed a "soft" solution using a single, unified hardware resource: an efficient, programmable multimedia processor designed from the ground up for the low-cost PC market. The result is a single chip that combines the performance to deliver concurrent operation among the seven multimedia functions, complete compatibility with existing PC applications (including DOS games), the flexibility to adapt to evolving standards and emerging technologies and the low cost to make it affordable as standard equipment on all personal computers.

While the Mpact media engine is low cost, it's performance is high power. Under the control of the Mpact media engine's real-time, multithreaded, multitasking operating system, multiple, highly optimized "MediaWare" configuration modules developed by Chromatic operate concurrently to provide the specific multimedia functions. These include:

Additional features or new levels of technology are achieved through a simple software upgrade of the appropriate MediaWare module via floppy disk, CD-ROM or on-line service.

Cost-Effective Design

The Mpact media engine is designed to do one thing very well -- crunch multimedia data to perform the required multimedia functions in various combinations simultaneously. But to be viable for the PC mass market, it must do this very efficiently to keep costs down.

Chromatic has achieved these remarkable speeds very cost effectively through its unique coprocessor implementation. By maintaining a sharp focus on PC multimedia functions, designers were able to assume the Mpact media engine would always operate in the presence of a PC host processor and could make intelligent use of its capabilities. For example, the PC host processor is sufficient for all the floating point calculations needed to generate the triangle vertices and display list reversal for 3D graphics. The PC host processor can easily do all the huffman decoding for MPEG processing, too. As a result, Chromatic was able to eliminating these functions from the Mpact chip, greatly simplifying its design and significantly reducing its cost.

Finally, by using a single, high-powered processor to handle all multimedia functions, no function-specific circuitry is ever sitting idle taking up precious and costly silicon real estate if a particular function is not being used.

Performance and Bandwidth

Multimedia processing requires billions of operations per second. MPEG decoding requires 1 to 2 billion operations per second, 3D rendering requires 1 billion operations per second (for 1 million triangles per second) and MPEG-1 encoding requires tens of billions of operations per second. But there's a limit to how fast tasks can be performed sequentially.

To achieve real performance gains, a processor has to do more than one thing at a time. The breakthrough design of the Mpact media engine combines elements of supercomputer and digital signal processor (DSP) technologies -- including a very high-bandwidth, parallel architecture, very long instruction word (VLIW) and single-instruction/multiple-data (SIMD) execution and vector processing -- to achieve the blazing speed and tremendous throughput necessary to perform the seven multimedia functions in real time.

Data moves into and out of the engine simultaneously over five high-speed I/O buses at up to 500 megabytes per second (MB/s), the same speed as the original Cray supercomputer, ensuring that the Mpact media engine is never waiting to send or receive data between peripherals, host bus or RDRAM. At the same time, a 792-bit-wide internal data highway is moving up to 8 billion integers per second between hundreds of arithmetic logic units working in parallel. As a result, it is able to achieve 2 billion integer operations per second for most functions, and an amazing 30 billion integer operations per second for the critical motion estimation function used in video compression/decompression.

Inside the Mpact media engine, a programmable core performs all the multimedia data processing required to execute the various audio, video and graphics functions. This core is surrounded by interface controllers for each of the five high-speed external I/O buses -- Rambus DRAM (500 MB/s), PCI (50 MB/s), display (200 MB/s) , video (27 MB/s) and peripheral (2 MB/s) buses. A single multimedia memory buffer comprising 2 to 16 megabytes of 9-bit wide Rambus DRAM acts as the central store for all multimedia data and MediaWare modules. RDRAM provides the necessary bandwidth with a single-chip, low pin count solution. Using direct memory access (DMA), the interface controllers move data into and out of the RDRAM buffer on a microsecond-by-microsecond basis where it is made available to the processor core.

The supercomputer-like architecture of the processor core combines five groups of parallel processing units, a multiport SRAM vector register file, an instruction decoder and RDRAM interface. The 11 outputs from the processing unit groups and the SRAM read ports are each 72 bits wide, and are arranged in parallel to form the 792-bit results bus. The results bus ties these outputs back to the 72-bit-wide processing group inputs, SRAM write ports and instruction decoder input using a data crossbar arrangement, effectively giving any input access to any output. If a result is to be used in the next cycle for another ALU operation, it passes through the crossbar into the appropriate ALU. If it's not to be used in the next cycle, it passes into one of the write ports of the SRAM where it waits until it is needed, and then passes out a read port. Reading and writing to the results bus can occur concurrently. By making the SRAM multiported and high speed (4 billion integers per second), execution can occur directly from SRAM with zero wait states. Data moves between the SRAM and RDRAM over the 500-MB/s Rambus.

Parallel Processing

To make the most efficient use of this parallel architecture, the Mpact media engine employs a combination of VLIW, SIMD and vector processing. Instruction words are fetched on every clock cycle, and each instruction word contains two instructions or opcodes. Each opcode, in turn, can operate on 2 to 16 integers simultaneously, enabling as many as 32 integer operations to occur in parallel for each clock cycle. In addition, vector instructions can automatically perform multicycle operations on an array of operands, providing much greater efficiency than if these operation were performed using traditional program loops.

Besides general purpose instructions, several application specific instructions provide performance boosts without requiring application specific hardware, holding the lid on cost. These instructions are aimed at the performance sensitive "inner-loop" functions used in discreet cosine transforms, bitblt and adaptive filtering functions, among others. By taking advantage of these highly tuned loops, MediaWare functions can run twice as fast.

Chromatic's engineers carefully studied these performance sensitive inner loops to extract maximum performance from the processor. They looked for commonality among the functions, designed building blocks and data paths for them, built instruction sets giving access to these building blocks and then wrote the firmware. They then tested the results against their performance objectives before freezing the architecture. Throughout the process, the engineers looked for opportunities to balance the proportion of adders and multipliers, the data widths, and other parameters, yet tried to retain enough generality to maintain high performance for unanticipated applications.

Unique Business Model

Chromatic's unique business model can be thought of as the next step beyond the "fab-less" semiconductor company. Instead of manufacturing the Mpact media engine itself, Chromatic has licensed its design to two of the world's largest semiconductor manufactures, Toshiba and Lucky Goldstar. Not only does this arrangement keep costs down through the manufacturing cost efficiencies inherent in high-volume production, it also enables the Mpact media engine to be implemented using the latest 0.5- and 0.35-micron process technologies, further boosting performance while reducing costs.

Creating the Mpact media engine also required specialized hardware and software skills not commonly found in traditional semiconductor companies. Through its ability to offer equity interests in the company, chromatic was able to attract some of the most respected technical talents in the industry, including Wes Patterson, executive vice president and COO of Xilinx, Mike Farmwald, founder of Rambus, Stephen Purcell, chief architect at C-Cubed, Dave Holt, vice president of engineering at SHOgraphics and Darrell Burns, lab manager at HP's Computer Technology Lab.

Chromatic Research

Chromatic research was founded with a single, aggressive goal: To redefine the personal computer through affordable silicon technology that integrates all seven major multimedia functions. As the first company to achieve this goal and offer the only complete multimedia solution, Chromatic is leading the way in the evolution of the PC from information processor to personal communication and entertainment center. Chromatic Research is well positioned to assume a leading role in the industry as technological capabilities and market demand for multimedia computing converge.