You keep making the point that general-purpose microprocessors are being pushed aside by application-specific hardware devices, and now devices that are application-specific hardware and software (Mentor's reason for acquiring Microtec Research a couple of years ago). While it is true that application-specific hardware devices will always be invented to handle demanding applications, you fail to understand the dual forces that constantly erode advantages of hardware, supplanting it with software. First, general-purpose microprocessors double in speed every 18 months. Second, functions that can be programmed using general processors can be produced at zero marginal cost.
Rapidly increasing speed of microprocessors means that the capability of microprocessors continually expand to absorb applications previously considered beyond the capability of general processing. The zero marginal cost of production mandates migrating applications to software, where possible.
Casting these proclamations in stone, they take the form of a law:
The Law of General Processing: Any given level of application functionality will eventually succumb to general-purpose processing.
But I don't believe for one minute that everything will succumb to general-purpose processing, so we need another law to stop this from happening:
The Law of Specific Processing: There will always be new applications that justify application-specific hardware.
I agree with you that neural networks is an area that appears to beg for exotic hardware solutions, namely massively parallel processing, seemingly not the forte of WIND. Yet, once more, things are not always what they seem. Did you know that VxWorks plays an integral role in the world's fastest, parallel-processing computer built by Sandia Labs along with Intel using over 9,000 microprocessors? The jointly-developed "nano-OS" on the compute nodes is code named "Puma" at the labs, and "Cougar" at Intel. VxWorks is used to monitor each processor's contribution and helps control overall operation. Surely that must impress any fan of parallel processing such as yourself.
What you implied about VCSI, the example I mentioned, is dead wrong. Yes, VCSI uses special hardware. They were a major supplier to Dialogic for voice recognition hardware. To this day they sell voice recognition black boxes and cards. Indeed, two years ago you would have labeled VCSI mostly a hardware company. That has changed, and that's my point. VCSI now considers itself mostly a software company, which also shows in its revenues (less revenues but reduced cost of goods sold). Its sales to Dialogic are now mostly, or maybe even entirely, software. VCSI's recent announcement about functioning entirely within the Pentium processor reinforces the company's shift to software.
Why is VCSI migrating to software? Because their algorithms for voice recognition no longer requires specialized hardware, and therefore, by the Law of General Processing, the migration is mandated . (By the way, you are unkind to VCSI when you made the condescending remark, "Even if you use a specialized VCS Antares DSP card, it will still on recognize 50-60 words." These cards target telephony command and control, which traditionally have limited requirements, which is reflected in hardware designs. With the software solution, the size of vocabulary is limited solely by available memory. That includes thousands of recognizable words in about 45 different languages.)
Does this mean that all voice recognition now must be processed by general-purpose processors? No. Other, more exotic algorithms, perhaps based on more complex neural networks, may still require special hardware - for now. By the Law of Specific Processing, we know these leading-edge applications will always exist.
Think of general-purpose processors as a black hole, forever devouring physical material (hardware-specific applications) that falls within its gravitational influence, doomed never again to see the light of day. But since the universe if filled with unlimited complex applications, there are always new applications to be devoured. To suggest that a black hole (general-purpose processor) is under pressure from free floating objects (hardware-specific solutions) is nonsense. The exact opposite is always the case.
Also, nothing said above precludes VxWorks from extending beyond general-purpose processing and uniting in special ways with hardware to solve vertical applications. It did it in creating the I2O chip. And I suspect similar developments are likely to occur in the future. This is consistent with the Law of Specific Processing.
Mark, I referred to your frequent identification of specific hardware applications, as though they threaten WIND, as harping. Its harping because there is nothing more that needs to be said; we agree. There are lots of existing applications of this type, and many more to come. Anecdotal descriptions of these applications do not diminish the need for general-purpose processing, they simply add to the growing list of future applications VxWorks will handle.
As to my harping on ubiquitous computing and WIND, I confess that I am guilty as charged. My only excuse is that these two items are the subject matter of this thread.
I wondered when you would pick up on Jerry Fiddler's remark in Fortune magazine that he doesn't count on retaining his wealth, based on the history of tech companies. In all honesty, don't you think Jerry was just being unpretentious? What if you woke up one day to discover you were worth $100 million, soon to be a billionaire? The risk is not that you will loose your fortune; the risk is in being wealthy without hurting people you love.
I truly hope this helps.
Allen |
