CNET también está disponible en español.

Ir a español

Don't show this again


What makes a PC fast

A PC these days is defined less by the processor's raw speed and more by chip technologies competing to rev up other parts of the PC--but it can all be pretty confusing.

A personal computer these days is defined less by the processor's raw speed and more by chip technologies competing to rev up other parts of the PC--but it can all be pretty confusing.

As the performance of processors from Intel and Advanced Micro Devices break through the 600-MHz barrier, other chip technologies are stepping up to reduce bottlenecks or lift performance, leading to a mix of solutions that can be confounding to many PC users--but crucially important nonetheless.

The stakes are high too for companies like Rambus, Via Technologies, and ATI Technologies as they vie to establish their technologies as the next performance must-haves in a personal computer.

These technologies lead to some obscure questions for the PC buyer. Will your PC have a 66-MHz, 100-MHz, 133-MHz, or faster, bus? Will it come with synchronous DRAM or RDRAM? How large is the on-chip cache? Does the graphics technology come close to true 3D performance or is it really 2D clothed in 3D marketing hype?

Many consumers may not care about these issues, but they will be increasingly important. Most of these technologies are used now--or are going into personal computers soon--and can impact performance as much as any AMD or Intel processor.

Pokey bus
Some of the new technologies simply try to reduce speed impediments. Two of the most pivotal points of congestion are the system bus and main memory.

Slowdown in these two areas can be enough to make a user wonder if their Lamborgini-like PC with a 600-MHz Pentium III isn't chained to a 1966 Volkswagen bus. "It's like you're driving down the highway at high speed and you hit a big, deep puddle," said Nathan Brookwood, a principal at Insight 64, a Saratoga, California-based consultancy and formerly an executive at Micronics, a PC circuit board maker.

The problem begins at the system bus, a conduit for shuttling data between the processor and other components like PC's main memory--and the bottleneck that is closest to the processor.

While the processor races ahead at 500 MHz, the bus runs at a pokey 66 MHz for some PCs. This is the case for Intel's fastest Celeron chip and similarly for the 400-MHz Pentuim II processor used in notebooks. For faster chips, Intel, AMD, and others mitigated this somewhat by boosting the bus speed to 100 MHz. But Brookwood cautions that for most users the faster bus doesn't necessarily translate to a dramatic boost in system performance. "There's something like a 3 to 5 percent difference [between 66 and 100]," he said.

Nevertheless AMD may have the last word on the importance of bus speed. AMD's new Athlon processor, which is due to appear in systems over the next few weeks, has a 200-MHz bus. "Athlon is clearly testing ahead of the Pentium III [in performance]. How much of this has to do with the bus, we don't know yet," he said.

Indeed, some analysts see the upcoming AMD bus as a very positive development. "The [current], processor bus is contributing to the slowdown," said Mike Feibus, a principal at Mercury Research. He said that this can cause problems for certain kinds of data processing, particularly on a Pentium III processor running at 600 MHz with a bus running at one-sixth the speed.

"It's like bringing a six lane highway down to one lane--depending on what time of day you're driving," Feibus said referring to what kind of tasks the processor is working on. "If there's a lot of traffic it can be pretty bad."

Current speed limit 133-MHz?
This gridlock really begins to get interesting, however, when the data slams into the speed bump lurking today in the PC's main memory. Not only is this a technical concern but also a business battleground for companies like Via and Rambus.

"The memory is just a whole lot slower than the processor," said Brookwood who refers to the problem as "latency." Here, data sits and goes nowhere as the memory tries to catch up. This is particularly an issue with large server computers. Rambus's memory chip technology, though still unable to eliminate the latency problem, will go a long way toward speeding things up while technology from Via is more of an interim solution, according to analysts.

So, the options boil down for the most part to this: 100-MHz memory, 133-MHz memory, or Rambus memory, which can run at speeds ranging between 600 and 800 MHz. The former two technologies are referred to as SDRAM and the latter as RDRAM.

Today, the faster Pentium II processors and all Pentium III processors use 100-MHz SDRAM called PC-100. Via and others, along with a number of memory chipmakers, are pushing to get the 133-MHz speed memory into PCs. Rambus, meanwhile, is starting to deliver samples of its RDRAM chips but at "price premiums double or triple [standard memory]," according to Feibus.

RDRAM can result in dramatic speed improvements but this may not happen until processors begin to get faster, according to Brookwood. "It's very hard to assess [ Rambus] at the present time," because of lack of testing. Brookwood believes Rambus won't really make a big difference until processors get much faster than today's because Rambus chips can "scale" up much better to match the speeds of superfast processors as they get speedier over time.

Intel will support Rambus with new chip technology in September. In the meantime, the chip giant will explore the 133-MHz option but has yet to state anything definite about an interim technology before Rambus.

So, the upshot is that faster 133-MHz memory is likely to be the solution over the next 12 months or so for the mainstream PC market. Rambus will come later because of current supply problems, according to Feibus.

For bus and memory constraints, almost any solution would help, according to analysts. "[Bus and memory limitations] can be a real bottleneck for things like streaming video," Brookwood said, an application used increasingly on the Internet.

Intel and AMD solve some potential bus and main memory limitations by increasing the size and proximity, to the processor, of the cache, which is special memory able to handle data in a much speedier fashion than that found in the PC's main memory. (SDRAM, for example, is considered the main memory.) Intel, for instance, now integrates 256 kilobytes of cache memory right onto its processors. This is expected to increase over time, improving performance as a result.

Graphics is hot--and not a bottleneck
Graphics chips from companies such as ATI, S3, Nvidia, 3DFX, Matrox, and NeoMagic are a boon since they take a load of the main processor and boost performance rather than creating a bottleneck.

"Graphics is probably the one positive exception," said Feibus.

What makes graphics chips tick? Parallelism is the esoteric but short answer. "In summary, parallelism is certainly the most important element of graphics architecture," said Peter N. Glaskowsky, an analyst at Microprocessor Report.

"The best chips are the ones that have the most parallel[ism]." This refers to a chip's ability to handle a number of tasks at the same time via 'pipelines.' "Mainstream chips from Nvidia, 3DFX, and others are at two pipelines, going to four in their next-generation devices," he said.

But he also said that though graphics chips are generally fast, in some respects there isn't that much difference between many of them in the mainstream PC market. "There's no real technology advantage as such, but some chips are a little faster than others, and some companies have done a better job with their text-scrolling and other minor things that have a big effect on benchmarks."

He also said that the promotion of 3Dism in graphics chips is overdone. "Almost everything we call a '3D' graphics chip really just draws 2D graphics. The 3D-ness comes from the polygons that define the image to be drawn. They're defined in three dimensions but projected to two dimensions during rendering--just as a camera converts a 3D scene to a 2D film image. Even this data is really just two-dimensional: it defines only the surfaces of objects, not what's inside them."

He said there are exceptions such as Mitsubishi's VolumePro family of chips.