Right, Transmeta
just kicked a ball rolling down hill and a lot of people are sitting there
who are saying "TM3120 and TM5400. Yeah, nice processors. Not exactly Earth-shattering
though, right?"
Wrong, and perhaps even Transmeta are unaware of the profound impact. Perhaps not.
Transmeta have designed a hardware core that works on an entirely different view of computing logic. In the new processors, Transmeta have defined and implemented ways to:
Define the complex parallelising hardware in modern Pentium-like chips, and reduce it to a series of instructions that define how it will pass through the same logic as we currently use the silicon for. This allows the processor core itself to be designed in a way that optimizes the processing of this stream of instructions.
If you give it the logic to be a Pentium it will become a Pentium. This is using the same efficient parallel processor logic as the original, it's just defined differently. In software. And every time the code passes through the logic, it gets optimised further. The more you run it, the faster it gets.
To make it more efficient, the processor has an optimiser with a cache. It's about to get a little recursive here, so bear with me.
The optimiser looks at the parallel pipelining code, and removes the unnecessary steps from the pipeline. Although it takes a little while for the software to do this, far fewer instructions are executed - or clog the pipeline - and the performance you get from your core increases overall.
The optimiser also remembers a cache of pre-optimised instruction sequences that it has executed before, and obviously it uses them again.
The result, one darn' happy processor.
Now here's where current practice stops. The processors were designed for low power consumption but Linus Torvalds and friends are about to make the next move. People are saying "It's going to be a dash for speed!"
But how? The obvious step is to allow the optimiser to not only re-sequence the code, but also to give it the ability to modify a chunk of the hardware logic, optimising as it goes. Wow, an amazingly fast, self-optimising processor.
But why stop at giving the optimiser its own little playground? Why not give it access to all the logical components of the chip? Allow it to reprogram every unit of the chip by turning it into a big Field Programmable Gate Array (FPGA). By giving the new optimiser the rules to understand the reprogramming of an FPGA (theoretically we could start doing this today with either a simple FPGA or something like IBM's Blue Logic), we end up with a processor that will actually optimise the design of the chip and is basically a huge array of repeating logic units.
These logic units would be in a "field" of a given number of units. If we were to increase the size of this field, the optimiser would optimise itself to spread into the new parts of the field. We now have an infinitely-scalable processor which is self-optimising. A true meta-computer.
Now we're into processing heaven. A very simple chip design, which scales infinitely and a processor which automatically optimises everything. With a little software assistance, these optimisations can be copied, enlarged and reduced according to need.
If we reduce the size of the field, we optimise the processor down to a minimum of gates. The optimiser uses less gates and more RAM when the field is being used up - something which can of course be part of the optimiser's software logic. What will the minimum size be for a processor that works? We shall find out shortly after building the first "field"-based computer.
What emerges as we start increasing the size of the field remains to be seen, but we will probably find out just how smart a computer can get relatively quickly.
Now it just happens that HP are developing a molecular processor which consists basically of a field of self-ordering, molecular switches. The new Transmeta software will be the optimal product to start using on future molecular processor chips. The computing power and power frugality of these devices defies the understanding of current processor capabilities. HP expect delivery by 2008 and I'm starting to both salivate and become apprehensive.
When design tools fabricated from this technique are used in the budding molecular engineering industry, particularly in the growing areas of DNA computing and molecular design (much like drug design software today) a deeper understanding of molecular chemistry is likely to result in a more efficient production of both field processors and the development of some new molecular tools. Designing molecules will shortly become an automated process, software to design extremely large scale molecules - as big as a house - will optimise itself, and The Nanotechnology Era will have arrived.
Unlikey? I think not. Transmeta's chip fabrication is being handled by IBM. IBM are currently undertaking the world's largest parallel processing project called "Blue Gene". The purpose of this project is to make a computer capable of designing proteins of specific shapes and specific characteristics; effectively designing small molecular machines. It will surprise me immensely if they do not take advantage of the sheer power and self-optimising capabilities of Transmeta's design to do it.
Back to Launchpad
At least people
have seen this site?