The introduction of ATA-66 by the storage community has spawned much confusion among hardware enthusiasts and the pages that cater to them. Some simply look at the specification's increase in transfer rate from 33 MB/sec to 66 MB/sec and conclude that ATA-66 equipment must be faster. Others, noting that the transfer rate's of today's disks come nowhere near 66 MB/sec, do a cynical one-eighty, accusing the industry of delivering a useless, hype-filled spec. Neither one of these attitudes is truly appropriate.
Originally conceived nearly two years ago by Quantum Corp., ATA-66 has arrived in today's products en force. All current drives from every major manufacturer feature ATA-66 interfaces. Several add-on PCI cards add ATA-66 to legacy motherboards. Some motherboards feature an ATA-66 controller built-in. A few non-Intel chipsets natively support ATA-66. Intel themselves, late to the party, will introduce ATA-66 support into their upcoming 440JX chipset. What exactly does ATA-66 purport to bring to the table?
The most publicized advantage of ATA-66 is, of course, the increase in buffer-to-host transfer rates to 66 MB/sec. Secondly, ATA-66's improvements in timing margins and error correction enhance transfer reliability, bettering system stability and also increasing performance by reducing or eliminating the retries that result from a failed transfer.
Now then, what does increasing buffer-to-host transfer rates buy the user? When the system requests data from the hard drive not found in an operating system or controller-level cache, it goes next to the drive's buffer. If all the data needed by a given setup were somehow found in a disk's buffer, transfer rates would be significantly higher with the system running ATA-66 rather than ATA-33. This, of course, almost never happens. On the opposite spectrum, if all the data needed by programs were never in the buffer, transfer rates would not increase at all. The reason being, of course, that data would have to be retrieved bit by bit from the hard disk to the buffer, then from the buffer to the system. Platter-to-buffer rates would thus be the bottleneck. Improving buffer-to-host transfer rates in this case would be useless since the drive can't actually get data off the platters and into buffer at anywhere near the same speed.
Applications in the real world lay somewhere between these two extremes. In most cases, all needed data is not in the buffer. Conversely, there are very few cases where at least some of the data requested won't be present in cache. The mix in contemporary applications has been revealed already with changes measured from DMA mode 2 (16.6 MB/sec) to ATA-33 (33 MB/sec). Yet ATA-66's release, an upgrade along the same line as going from DMA mode 2 to ATA-33, nevertheless has stirred up much doubt and confusion.
Frankly, we're surprised that StorageReview.com needs to weigh in with our take on the matter. Back when ATA-66 initially hit the marketplace, we were content to stay away from an "ATA-66" article; our hands were full with what we felt were more important matters, such as reviewing the latest drives just starting to trickle out. We believed that a simple thesis (The disk read mix found within the vast majority of applications is dependent on disk-to-buffer transfers rather buffer-to-host transfers; thus, there will be little change in application-level benchmarks when going from ATA-33 to ATA-66 operation) would be proven by the inevitable tests that would roll out of the "everything under the sun" hardware web pages.
This has not been the case. Unfortunately, results turned in by various hardware pages and enthusiasts on the subject have varied in data and conclusions. In almost every case though we can easily find fallacies, over-looked controls, or other experimental errors yielding faulty data. Let's examine some of these, and what StorageReview.com has done differently to ensure proper, controlled testing.
1. "Our tests show little performance increase going from ATA-33 to ATA-66. However, we used a 5400rpm drive for the test. A 7200rpm drive may benefit from ATA-66 operation." Articles like these -almost- hit the nail in the head, but discredit themselves with such a caveat. A statement along these lines displays the writer's ignorance when it comes to spindle speed and its relationship with transfer rate. The "66 MB/sec" in ATA-66 affects transfer rates. If a drive posts sequential transfer rates (i.e. disk-to-buffer rates) of 20 MB/sec, it's not limited by a 33 MB/sec. It's not limited by a 66 MB/sec limit either. Elementary! Even so, to minimize criticism from those who may buy the argument quoted above, our tests were performed on the Quantum Fireball Plus KX, at the time of this article the fastest ATA drive in both high-level and sequential transfer rate benchmarks.
2. "The drive posts substantial gains in transfer rate and high-level performance when run off of the [Promise/Abit/etc.] card instead of the motherboard's ATA-33 controllers. Further, ATA-66 mode uses dramatically less of the CPU's time than ATA-33." There are a couple problems with articles containing any of these statements. First, running the drive off of a legacy controller on a motherboard and comparing it to an add-in PCI controller, ATA-33 and ATA-66 not withstanding, is predominately a test between the controllers themselves, their firmware and their drivers. Differences between ATA-33 and ATA-66 simply get lost in there somewhere. Second, most add-in disk controller drivers enable busmastering by default. Windows 9x, however, does not default to busmastering. If the reviewers (and there've been quite a few) are unaware of such parameters, "ATA-33 vs. ATA-66" becomes "PIO mode 4 vs. ATA-66." A dead giveaway of such an error is a conclusion lauding ATA-66's reduced CPU utilization. In our investigation, both ATA-33 and ATA-66 tests were run off of the Promise Ultra66, thus eliminating controller/driver differences and ensuring identical busmastering techniques in both cases. ATA-33 was forced by disabling ATA-66 operation on the Fireball Plus KX.
3. "While we used an 80-wire cable to adhere to specifications when running the drive in ATA-66 mode, we used a 40-wire cable to force the drive to run in ATA-33 mode off of the ATA-66 controller." 80-wire cables reduce cross-talk and thus transmission retries. Thus, if "ATA-33" encounters enough cross-talk that retries start to mount up, there could be significant differences in speed from "ATA-66" operation. In both ATA-33 and ATA-66 operation, we used an ATAS Ultracable. ATA-33 operation was forced by disabling ATA-66 operation with a utility provided by the manufacturer rather than by changing cables.
4. "My brand-spanking new ATA-66 drive creams my three-year old ATA-33 drive." Comparing, say, the Quantum Fireball Plus KX (7200rpm, <8.5 ms seek time, 6.8 GB/platter, 512k buffer- and ATA-66) with the Quantum Fireball ST (5400rpm, 9.5 ms seek time, 1.7 GB/platter, 128k buffer- and ATA-33) to ascertain the differences between ATA-33 and ATA-66 is laughable. Nevertheless, "proofs" that ATA-66 is indeed faster with such evidence are disturbingly common. We used the same drive in both tests. Any questions?