The prior Radysis post on this thread reminded me of the fact that in order to understand what makes this era the digital age so enigmatic and at the same time one which offers so much opportunity. Those with the idea of the kinds of applications that will be utilized in the future relative to this point in time. As the core co's TI ADI Mot LU Siemens have begun to develop powerful fasterr chips with more on chip memory more in and off flow speed at the chip/ bus interface it has brought about the possibility to have board companies like Spectrum innovate/ These innovations are what they do. They can pile multiple chips on a single board and share memory between chips increasing the degree of integration of the board reducing cost [fewer chips/ less power etc. When it comes to highly complex dsp boards new innovative architectures must be developed that allow for these effieciencies. Going to the next level requires integrating multiple boards all operating as a unit ie inter board architecture and bus interface become issues. When you read the papers presented by the engineers as Spectrum along with those from academic institutes you get an idea of just how complex and sophisticated this integration is. It requires a thorough understanding of the dsp core, the board, and what needs to be added to make it perform maximally. This often requires development of new algorithms, new perceptions such as shared memory, and new ASICs' Application Specific Integrated Circuits. These chips are the glue of the board that a;;pw them tp function at peak performance for a given application with the memory and power crunching on board. That is why TI and ADI bring board vendors notabley the largest and arguably the best board integrator is in Burnaby. When new silica is under design power speed heat energy on board memory I/O, chip bus type all become important. The function of a dsp chip is simple inidea. Take the real world; speed, motion, temperature, altitutde, attitude, what ever and convert it, filter it, and make it speak in the language of computers through a translator equation or algorithm. Once this conversion occurs the chip can manipulate store share or act on the signal. It can then be progammed to send a signal to the board that can illict a reaction to the real world manipulate it in real time. In order to accomplish this the board must have an on board system that allows this massive number crunching ability to be coordinated with the function of other chips as part of the board level system./ Enter 3L diamond and Alex Apex RTOS {real time operating system} that acts like a traffic cop on the board making sure that the , timing of the boards components is accurate dependable and efficient.
Next in terms of increasing the power of a dsp system is when multiple boards need to be chained together to provide massive parallel power.
Who does this kind of innovation?People with vision, strong mathematics background, strong electrical engineering background, an broad understanding of architectural design; ie artisans. At Spectrum they have some of the most talented dsp engiuneers in the world bar none. The company like PMCS has been awarded the outstanding tech co of year in BC. They have been raised to tierI at TI. They have received honor from Siemens and Kodak as certified suppliers a rare distintion. They have been recognized in the wireless arena, motor contro,l radar, sonar, and in the semi industry. They are the big boys of system board levl developers. While the industry is just beginning to mature Spectrum has been fortunate to have Barry Jinks the visionary who has recognized abd attracted the talent to Burnaby and created a place where they could thrive. They had the fortune to attract Ron Wages the young star marketing guru from Texas Instruments. They have been smart enough to recognize the value and desirability of having a software division within the company for tooling and software are what separate the board companies. They have recognized the value and distinction of cores offered by TI and ADI. In the near future the demand for dsp boards will explode. Why because they add features improve quality reduce cost and are just simply better than analog systems. The digital age is everywhere in commercial appliances, autoelectronics, etc.. Soon they will be running factories contolling traffic flows/ and allowing for digital communications in secure broadband modes wireline and wireless. The age of connectivity of person to person person; machine to machine and machine to person..... to totally change the way we live.
I took the liberty of putting a couple of briefs from last year in which Spectrum engineers discussed some of the innovations they were developing.
Distributed Shared Memory - A New High Performance DSP Board
Architecture
Author(s): Mohammad Darwish, Frank Van Hooft, Spectrum Signal
Processing, Inc., Canada
Application Area: Parallel Processing
Summary: A new DSP-optimized single-chip DSP-PCI bridge chip has allowed the development of "Distributed Shared Memory", a new high-performance DSP board architecture created to fully utilize the ŒC6x¹s processing power. This paper discusses the architecture and its capabilities, such as the elimination of global shared memory, doubling the data transfer rate.
Abstract: A new single chip DSP-PCI bridge chip optimized for DSPs has
resulted in the development of an innovative high performance DSP board architecture called "Distributed Shared Memory".
The development of the Texas Instruments TMS320C6201 (ŒC6x) DSP has
challenged DSP designers to create new board architectures. New
architectures are required so that the tremendous power of the ŒC6x can be fully utilized. In general, traditional DSPs architectures
essentially have more I/O bandwidth than capabilities to process the
data. However, due to processing capabilities of the C6x, these
traditional architectures would result in the C6x processors being idle for certain periods of time while they wait for data. This is clearly inefficient. Therefore, to exploit the potential of the C6x, it is incumbent upon DSP systems designers to create new DSP systems
architectures. One new architecture is Distributed Shared Memory.
The Hurricane DSP-PCI bridge chip was designed specifically to work with DSPs. Consequently, it offers significantly higher performance than traditional DSP bridge chips. Furthermore, it replaces a voluminous two chip solution with one chip in a relatively small package. These capabilities, together with Hurricane¹s 64 word FIFO and DMA engine, has allowed the development of the Distributed Shared Memory architecture.Distributed Shared Memory is a DSP board architecture that eliminates the requirements for global shared memory. Therefore, with Distributed Shared Memory:
1. data can be transferred directly in and out of each C6x SBSRAM bank
from PMC and host via Hurricane; and
2. each DSP can read the SBSRAM of all other DSPs via Hurricane.
Distributed Shared Memory also provided dual or quad ported memory for
low latency message passing and interleaved access to each C6x¹s
external memory via the host port.
In traditional DSP board designs, global shared memory has been used as interim storage prior to placing data into the target DSP¹s dedicated memory. Interim storage means that data is handled twice; once to move the data into global shared memory, and once to move the data to the target memory. Eliminating global shared memory means that data only has to be handled once. Therefore, all things being equal, Distributed Shared Memory essentially doubles the rate with which data is moved onto the board or between processors.The result of the Distributed Shared Memory architecture is that processing potential of the ŒC6x can be fully utilized. The result is substantially higher hardware densities than would result from traditional DSP system architectures.
System Level Issues and Solutions for Multiple Board, Multi-Processor
DSP Systems
Author(s): Albert Lo, Spectrum Signal Processing, Inc., Canada
Application Area: DSP Systems Session 2
Summary: Designing multi-processor, multi-board DSP systems presents
challenges that are uniquely different from designing a single DSP
board. For example, integrating two different kinds of DSP boards into
one environment presents challenges that do not need to be considered
when using the DSP board in a homogenous environment. In this paper, a
method of addressing some of these system level DSP hardware design
issues will be presented and expounded. The method described here is
called the System Integration method. Where it was used, the System
Integration method has contributed to the improved quality of Spectrum
DSP boards.
Abstract: In designing multi-processor DSP Systems in today¹s world, the designer has many choices to consider. The main factors would be the choice of DSP to use (ADI, Motorola, or TI), and the type of system bus to use (PCI, cPCI, VME, or VXI). Possible system level design issues which should at some point in time be considered when designing a robust multi-processor DSP system include:
1) DSP Bus Arbitration - In designing the multi-processor DSP board, the designer usually needs to consider an arbitration protocol for the DSP bus. It is known that data throughput speeds is affected by the bus contention of the different processors sharing the DSP bus. What is not known, or is hard to characterize, is how the Bus Arbitration protocol will perform under stress, and in a multi-processor environment. Also, uncovering bus deadlock is difficult and only dealt with at the board level. If not considered, this can become costly when trying to build a multi-board, multi-processor DSP system.
2) Integrating different DSP boards together -
Most multi-processor DSP boards use either a Cluster Bus Architecture,
or a Mesh type Bus Architecture. In either case, there is a DSP
Processor Common Bus embedded in the board. Interfacing this DSP bus to the system bus (PCI, or VME) presents a few challenges to the designer. This issue especially becomes prevalent when inbound and outbound transactions between the DSP bus and the system bus occurs at the same time. These system level design issues are usually not considered when designing and testing a single DSP board. This is because the above two issues typically manifest themselves when a multiple board,
multi-processor DSP system is performing under stress, and system level factors such as bus acquisition times, bus propagation delays, become prevalent. Therefore, they cannot be ignored when building a
multi-board, multi-processor DSP System.To address these issues, a method of system integration was developed.The System Integration Test Method is a set of programs which will rigourously test most of the resources on a newly designed DSP board at the board level and at the system level. The System Integration test method, or SYSINT, will identify any errors in the DSP Bus Arbitration scheme, and provides a reliable method of ensuring that the timing, and compatibility requirements are met for a multi-processor, multi-board
DSP system before it is shipped to the customer.In a multi-processor, multi-board DSP system, the typical problems that occur can be traced to:
a) inability to handle inbound and outbound transactions under stress
b) inability to meet tight timing requirements of the hardware state
machines which would lead to missed transactions.It is during thecourse of data transfers between DSP¹s, or between DSP¹s and shared resources that these system level errors manifest themselves. SYSINT attempts to simulate these data transactions using software methods, rather than using different physical hardware, and thereby identify these system level errors. This is done in three ways:
a) design specific, and strategic read/write tests between DSP¹s, or
between the DSP and shared resources
b) introduce a degree of randomness
c) simulate as many different bus and timing cycles as possible.
SYSINT is modular in nature, and thus provides the advantage that it can be used at both the board level, in identifying any problems in the DSP Bus Arbitration, and at the system level, in identifying any timing, or data transfer problems between boards.
This paper will describe in further detail:
a) the system level design issues to consider when designing
multi-processor DSP systems,
b) the System Integration Test Method as a technique to address these
issues, and
c) a case example using Spectrum¹s Morroco SHARC-based DSP VME Carrier
Board.
The bottom line you don't have to be an engineer to enjoy the benefits of the digiatl age just sit back and worder How'd they do that?
Regards
Norden |
