|
|
|
|
Issue 23 Occasionally, a number of technical advances come together to give a quantum leap forward. This occurred recently as a result of three factors - the increased density of components on an integrated circuit; the successful fabrication of fault-tolerant complete integrated circuit wafers; and a new approach to structuring these wafers patented by Catt called the Kernel Logic invention. The result is that the latent, explosive power of semiconductor technology can be unleashed - one million computers working together in an array to solve large, complex problems at high speed. The technology is proven by the successful introduction into the U.S. market in 1989 of a solid state memory called "Wafer Stack", based on an earlier Catt patent. This received the "Product of the Year Award" from the U.S. journal Electronic Products in January 1990. (Details are in Electronic Design, 26th October 1989, and featured on the cover.) However, "Ivor Catt has always insisted that memory products are merely an incidental spin-off from the main work of wafer-scale development. The main purpose of wafer-scale technology, he believes, is to assist in the design of systems that will revolutionise computer architecture." - New Scientist, 12th June 1986, page 36. "WSI memory, says Catt, is a stepping stone towards the major exploitation goal of WSI, the design of massively parallel computer systems." - Parallelogram International, August 1990, p9. We can now advance to a more powerful implementation of the same technology. This is the array processor called The Kernel Machine. It contains one million processors giving one million million operations per second (one teraflop). It is built from an array of 100 wafers, and retails for £500,000. The external controller draws a map of good and bad chips, and devises a strategy for building a volatile, square two dimensional array of 1,000,000 perfect processing elements (PE's) out of a larger, imperfect array.
A general description of the machine is contained in Electronics and Wireless World, March 1989, page 254, and more briefly in the Sunday Times Innovation Section, 12th March 1989, page D 14. Worldwide patents, called "The Kernel Patent", owned by Catt, prohibit competition. Details are available from Catt's patent agent Graham Coles, tel 049 467 7181. Pats U.S. no. 5,055,774, Oct. 8, 1991; Europe no. 88905982.0 The Japanese patent application (no. 63-505902) is proceeding satisfactorily. The Kernel Project puts four working machines into service in the field. Summary of project cost. Each aspect of the project is based on access to one expert in that field, on whose judgement reliance is placed during the costing and project planning phase. Initial planning programme. Ivor Catt. This occurs during the six months before start of the project proper and costs £100,000.
Total Project Cost £20.7M
After an initial pre-start planning project of six months costing £100,000, the main project proceeds as follows; Start at a spend rate of £x p.m., rising
linearly to £3x p.m. at month 12.
The months of the project are named m01 to m48. In m42 the first two machines become operational, one in the field and one with a customer. Three more machines become operational on customer sites in m48. Main activity areas.
Critical paths. Portions of the project can run in parallel, so the critical path is not overly critical. However, a more hasty project would generate unnecessary problems. There are two main critical paths;
The project includes a great deal of simulation. Simulation will be of restricted value because of the massive scale of the machine, which is without compare. Reduced versions of programs will be needed which can be run in reasonable time when simulated on the less powerful machines which exist today. Our inability to simulate a machine with the power of Kernel indicates evolutionary advance. More powerful modes of operation of the Kernel Machine will have to be simulated on a Kernel Machine operating in a more primitive mode, rather than on the inadequate alternative machines. Our machine will initially have a primitive processor in each processing node, with space to accomodate the later upgrade(s). Alternatively, there will be a switch which will put either the primitive or the upgrade processor into use. Simulation software (Stage 1) will be delivered to customer site 18 months before the machine becomes operational on site. It will be important to get a machine operating in the field as soon as possible, with limited capability. A number of stages of simulation will be necessary.
The project is well balanced, being distributed among a number of cost centres. This means that deviations from the plan will minimally affect the total cost. The WSI development, at £0.5M, is surprisingly low since that would initially seem to be the major activity with the major complexity. This low cost results from access to Graham Davies and Mike Brent, who have gone through the technically critical phases in earlier Anamartic projects. However, at £0.75M, liquid cooling is larger than might be expected. Also, software development is a large item. This latter is fragmented into different types of software which do not interact. Again, deviations from the plan will not overly affect total cost. Project overview. Portions of the project tend to run in parallel, so the critical path is not overly critical. Each aspect of the project is based on access to one expert in that field, on whose judgement reliance is placed during the costing and project planning phase.
It is clear that an initial project planning exercise costing £100,000 is needed before commitment to the full project. This includes liaison with potential partners like GE and British Gas. It also includes the selection of the initial application. Long term break-even is achieved by selling into one application area. Profit comes from exploitation of further application areas. Examples of applications are; CFD (Computational Fluid Dynamics); body scanning; weather forecasting; air traffic control. The strategy is to give a monopoly to one company in the first chosen application and so expedite its rapid exploitation, while retaining full freedom in all other application areas. |
