NeuroMatrix® Architecture Is Chosen as a Part of the Best DSP*
Third-Party Program in the Industry
MOSCOW, Russia January 28, 1999 - MODULE Research Center, the owner of
NeuroMatrix® DSP* architecture (patent pending) joints Texas Instruments
"Customer Choice Third-Party Network" (www.ti-dsp.com).
The TIs Third-Party Program is designed to assist about 170
TMS320 third parties worldwide in gaining exposure for their products
and service and helping them align with TI business strategies.
RC "Module" has been established in 1990 by well known Russian firms of
military-industrial complex - Interstate corporation "Vympel" and its
research institute for radio devices (RIRD), entered to the program with
its new product - NM6403 processor. RC "Module" is a rapidly progressing
company known among the leading Russian and overpasses' companies
engaged in the design and development of embedded computer systems based
on modern electronic components and technologies. RC "Module" has a
staff of highly qualified engineers and scientists using the most modern
design tools such as Cadence and Synopsys and prototype manufacturing
equipment, which permits a very short design cycle (few months) from the
concept or idea to the finished prototype ready for mass production.
Although RC "Module" was founded nine yeas ago, the microprocessor
division was established in 1995. The first goal of the division was
designing the architecture and chip for best support of original neural
network and image processing algorithms developed by RC "Module". Due to
the usage of TI TMS320C40 DSP* for accelerating boards since 1994, there
were a lot of application software designed and optimized for TIs DSP*
architecture. That was one of important thing to design the processor
with communication capability to TI C40 DSP*. The NeuroMatrix® processor
architecture was developed in 1996. It includes 32-bit RISC* core, 64-bit
VECTOR core and two communication ports hardware compatible with TI
TMS320C4x. The NM6403 DSP* is a first hardware implementation of the
architecture produced using SAMSUNG 0.5-um CMOS* technology appeared in
1998. The main feature of NM6403 processor is a programmable operand
width and scaleable performance from 50 MMAC (32-bit data) up to 51.2
GMAC (1-bit data) at 50 MHz clock rate. The communication ports allow to
use NM6403 as a vector-matrix co-processor for multiprocessor systems
based on TI C4x DSP*. The processor is able to perform a broad range of
DSP* applications. Thus, the 256-point 32-bit complex Fast-Fourier
Transform (FFT*) performs in 4070 clock cycles and require only 80
microseconds, the Walsh-Hadamard Transform (WHT) with 2M points
calculates for 0.34 seconds, the Forward Propagation (1024 layers, 1024
neurons/layer) neural network emulates for 1.53 seconds, "Sobel" 2D
filter performs 37.5 frame(384x288 pixels)/second.
NM1 Dual-CPU is a first PCI* board based on NM6403 DSP*. The board
contains two NM6403 processors, up to 8 Mbyte static memory, 64 Mbyte
dynamic memory. The performance of NM1 for vector-matrix calculations is
1.9 billion operations per second. The board has four communication
links which allow easy connection to any Third-Party DSP* boards based on
TI C4x. The development tools for PC host platform includes an ANSI
X3J16/95-0029 preliminary standard compatible C++ compiler, an
assembler, an instruction level simulator, a cycle accurate simulator, a
linker, a source level debugger, a load/exchange library and set of
application specific vector-matrix libraries. The vector-matrix
optimized library simplifies C-language programming and allow to design
the DSP* and telecommunication applications. The NM6403 Verilog silicon
proven model and test bench is also provided for system level simulation
and PCB* design.
The Single-CPU one-size TIM (Texas Instruments Module) based on NM6403
DSP* will be available on July 1999.
Module® and NeuroMatrix® are registered trademarks of MODULE Research
Center. All other trademarks are the exclusive property of their
respective owners.
www.ednmag.com/reg/1999/041599/08cs.cfm - Article in EDN