Campbell: Processors Repeating Mistakes

The computer world took some wrong turns in its epic trudge toward the most powerful processor.

Luckily, those wrong turns have generated some of the most brilliant advancements ever made in processor design.

The first processors, which were generated using a whole bunch of individual chips, were found to cap quickly, leading to the brilliant idea of putting millions of transistors on a single chip die.

Intel’s attempt to clock the hell out of its Pentium 4 processors, generating one in excess of 4.0 GHz, resulted in a chip that would easily melt most computer cases and bankrupt most municipalities in power consumption.

This setback, however, led someone to speculate that a chip with multiple, somewhat independent processor cores could provide just as much horsepower without shutting down the power grid.

Its journey took a ridiculous turn with a simple thought: If having two cores is good, then having hundreds of cores is better.

According to the online newsletter GoodGearGuide, Tilera, a semiconductor design company, announced the release of the Tile-GX processor, the first general purpose processor to utilize 100 independent processor cores.

At face value, this seems wonderful.

A Dell PowerEdge server running one quad-core processor is extremely fast. A computer, theoretically, running the Tilera processor with 100 cores, would be 96 cores faster.

There are problems, however.

First of all, the Tilera processor is somewhat power hungry. At full capacity, it requires 55 watts of power.

Although this is not much when compared to the single-cored Pentium 4s – which could suck up around 65 watts — it is a bit more than the Tilera 16-core processor that could use around five watts of power.

It becomes a rather confusing gray area to consider whether the processing power and parallelization afforded by 100 cores is worth the significant power reduction over a still-formidable collection of 16 cores.

The correlation between the number of cores and performance is also somewhat confusing. It is not as easy as “the more cores, the more power.”

Consider an office building full of workers. If there are five workers, each worker is more likely to be efficiently utilizing his or her time, and each worker is much simpler to control from a centralized boss.

The boss can keep each worker working efficiently, and as an end result, a large problem can be easily solved.

Now, have that company expand and hire 95 more workers. To remain as efficient, the office needs to streamline and expand to handle the larger capacity.

The communication channels between each employee — the phone system, for example — must be expanded to handle extra traffic. More managers must be hired to effectively manage the army of new employees.

Despite an increase in manpower, the office might remain just as efficient or, in circumstances of bad management, lose efficiency.

This is not to say that many core systems are not a thoroughly bad idea.

Graphics processors utilized tens or hundreds of cores before. They are beginning to make their way into general purpose computing, but they are experiencing snags.

Large numbers of cores are excellent at crunching numbers.

They are not necessarily as efficient at performing the constant I/O operations that characterize everyday computing.

The Tilera processor might find 100 cores inefficient at performing file reads or efficient memory allocation, among other tasks.

Tilera published accounts of its new Tile-GX processors, claiming that its configuration of grids of processors is superior to the multi-core technology currently being championed by Intel and AMD.

The grid pattern, Tilera claims, makes its processors more easily scalable and allows for more efficient communication between processor cores.

If true, the Tilera processors might indeed overcome some of the efficiency problems that characterize multi-core systems.

Tilera also championed backward compatibility and allowed its processors to communicate with Intel x86 processors connected in parallel on a computer’s motherboard.

This communication could enable the Tilera processor to handle the problems it can solve most readily, while leaving more difficult problems to an Intel processor with fewer cores and fewer complications.

Ultimately, as processing demand outstrips the capabilities for even the most powerful of current multi-core processors, some breakthrough must be made in making the division of computer problems and the channels through which cores communicate more efficient.

Tilera might very well have made this breakthrough.