Section 5.1. The Truth About Processor Performance

5.1. The Truth About Processor Performance

Processor companies do nothing to discourage longstanding myths about processor performance. It's true that in the early days of microprocessors, a new model was often two or even three times faster than the model it replaced and sold for little or no more. In those halcyon days, the fastest available processors were sometimes 10 times faster than less expensive models that were still being sold.

There was also a favorable bang-for-the-buck ratio. If you paid twice as much for a processor, it was probably considerably more than twice as fast. We remember testing our 4.77 MHz IBM PC/XT against a 16 MHz 286 PC/AT when both were still being sold. The latter system cost two or three times as much, but was something like 10 times faster.

Those days are long past. Nowadays, processor performance increases incrementally, the accompanying price differences are large, the performance gap between the slowest and fastest current models has narrowed substantially, there are many, many more intermediate models available with minor performance differences, and the bang-for-the-buck ratio for the fastest processors has dropped well below 1:1. AMD and Intel have both learned to "work the market," maximizing their revenue in a very competitive market.

5.1.1. Price and performance

Here's a dirty little secret that AMD and Intel would rather you not know. At any given time, the actual performance differences between their slowest and least expensive "economy" processors and their fastest and most expensive "performance" processors is relatively small. A $750 processor you can buy today will probably be at most 2.5 to 3 times faster than the $50 processor sitting next to it on the store shelf.

Doubling or tripling performance may sound like a huge improvement, but human perception is not linear. A processor must be 30% to 50% faster than another processor before most people perceive any noticeable difference in routine use. Doubling processor speed results in an obvious difference in performance, but not a knock-your-socks-off change. Tripling processor speed provides a very noticeable performance boost, but at a very high price.

And most of that performance increase comes at the lower end of the price continuum. Paying more for a processor yields rapidly diminishing returns. For example, a $175 processor may be twice as fast as a $50 model, or nearly so. Doubling the price to $350 may buy you only a 25% faster processor, and doubling the price again to $700 less than a 10% bump in speed.

ADVICE FROM RON MORSE

Remember, processor speed is only one of the things that determines perceived system performance. Memory speed, bus speed, hard disk performance, and video performance all play a role, albeit to a lesser degree. Still, the fastest processor won't much help a system that's I/O-bound or has a slow video adapter.

But all this is true only for a particular moment in time. As AMD and Intel discontinue older processor models and introduce new ones, the whole continuum of processor performance shifts upward in lockstep. A midrange processor today is faster than the fastest performance processor of a year or 18 months previous, and even today's inexpensive "economy" processor is faster than the fastest processor of 2 to 3 years ago. That's good news, because it means it's often possible to upgrade an older system to today's level of performance at a reasonable low cost.

SOMETIMES EXPENSIVE COSTS LESS

None of this is to say that there's no place for very expensive processors. For programmers who spend their days doing iterative compiles and links, saving just a few seconds on each iteration may well be worth the higher cost of a premium processor. The same is true for highly paid executives, for whom lost seconds translate to lost dollars; for commodity traders, for whom seconds may mean the difference between a huge profit and an equally huge loss; and for serious gamers, who need every advantage they can get (not to mention bragging rights).

5.1.2. AMD versus Intel

Fanboys and brand zealots argue that AMD is faster than Intel, or that Intel is faster than AMD. They're both wrong, and both right. The truth is that at any given price point, Intel and AMD processors are remarkably closely matched in overall performance. That's not to say that their performance is identical for every application. AMD processors, for example, typically have better gaming performance than similarly priced Intel models, and Intel processors typically have better multimedia performance than similarly priced AMD models.

THE PROCESSOR TUG-OF-WAR

For a time, Intel processors badly beat AMD processors by most performance measures. AMD countered this fact by improving their own price/performance ratio, increasing the speed and cache size of their processors, and reducing their prices until they were again competitive with Intel processors that sold for similar prices. Then, for a time, AMD processors badly beat Intel processors by most performance measures. Intel countered this by improving their own price/ performance ratio, increasing the speed and cache size of their processors, and reducing their prices until they were again competitive with AMD processors that sold for similar prices. Each time a gap in price/performance occurs, Intel or AMD adjusts processor pricing and/or performance levels to eliminate that gap. Both companies are smart enough to keep the playing field level while not leaving money on the table.

5.1.3. Benchmarks lie

Benchmarks are supposed to provide neutral measures of the performance of processors, both overall and in terms of specific types of tasks. But modern processors are very complex devices, with numerous strengths and weaknesses relative to competing processor models. A benchmark test that happens to play to a strength of a given processor will make that processor look (unjustifiably) very, very good. Conversely, a benchmark that gives heavy weight to a function that happens to be a weak point of a particular processor will make that processor look very, very bad, again unjustifiably.

If you allow us to choose the benchmark tests, for example, we can "prove" that a $150 AMD processor is faster than a $1,000 Intel processor. But by choosing different benchmark tests, we can just as easily "prove" that a $150 Intel processor is faster than a $1,000 AMD processor.

Broadly speaking, there are two types of benchmarks. Synthetic benchmarks are designed to test individual aspects of a processor's performance, such as cache efficiency, memory throughput, or floating-point performance. Application benchmarks, also called natural benchmarks, incorporate several common applicationssuch as MS Word, Adobe Photoshop, LightWave, and so onwith predefined suites of tasks to be performed.

The knock on synthetic benchmarks has always been that they are "meaningless" because they don't measure real-world task performance. We think that's wrong-headed. For example, if we want to decide which processor is likely to be fastest for applications that are bound by memory performance, we can use synthetic benchmarks to test the memory throughput of different processors. The results of those synthetic benchmarks will in fact give us a very good idea of the likely relative performance characteristics of different processors.

Conversely, using application benchmarks provides useful information only if we happen to be running the same applications used in the benchmark suite, in the same way, and with the same relative weighting. Two processors might achieve very similar overall results in an application benchmark, and yet one processor might be a better choice for running one of the suite applications, while the other processor might have the edge for running another.

5.1.4. Optimizing price/performance ratio

Dollar for dollar, AMD and Intel processors typically have very similar overall performance. We follow a few simple rules when we choose a processor, and suggest you do the same:

• At the low end, $50 to $125, AMD processors dollar for dollar provide noticeably better performance than Intel processors across the board. Intel has always paid lip service to the low-end market, but it really has no interest in competing here. It costs Intel about $40 to make a processorany processorso it prefers to devote its efforts to market segments with higher profit margins. On the other hand, the low-end market has until recently been AMD's bread and butter, so they devote a lot of attention to this segment.

• Choose a low-end processor unless you have a good reason for spending more. For most older systems, the most cost-effective upgrade is a processor that sells for $50 to $75, whether your motherboard is AMD- or Intel-compatible. Low-end processors are perfectly suitable for most computing tasks, including productivity applications, web browsing, email, watching videos, and so on.

• In the mainstream range, $125 to $250, Intel and AMD processors are pretty evenly matched dollar for dollar overall. Profit margins are much higher here than in the low-end segment, and unit volumes are huge, so the competition between AMD and Intel here is fierce.

• If you put heavier demands on a processor, such as casual video editing or 3D gaming, spending an extra $75 to $125 on your processor upgrade can provide major benefits. Processors in this price range are typically noticeably faster than low-end models, and for some applications that additional performance matters.

• At the high end, $250 to $1,000, Intel processors are generally somewhat faster dollar for dollar than AMD processors overall, particularly dual-core models. Although AMD produces the fastest processors in this segment, they set very high prices for those processors compared to Intel models that are only marginally slower, so Intel wins the bang-for-the-buck competition. Profit margins here are very high, but unit volumes are very low, so the actual dollars at stake are much less important than those in the mainstream segment. AMD and Intel compete here mainly for prestige and bragging rights.

• Unless the current system is very recenta year old or lessit almost never makes sense to upgrade to a high-end processor. The potential incremental performance benefits of a high-end processor, limited as they are under optimum conditions, are even further limited by the low performance of other components in an older system. Furthermore, it's probable that installing a high-end processor would also require that the motherboard, power supply, and possibly the memory be replaced, which amounts to building an entirely new system. The one exception here is devoted gamers, some of whom think nothing of installing a new $1,000 processor every six months (not to mention a new $700 video card, or two.)