Identifying Problems In PC Hard Disk Recording Systems

"Hard disk recording is very unlike most mainstream applications in the way that it stresses the PC."

All computers are put together using a set of components, but in the case of PCs the components can come from a wide range of manufacturers, with an equally wide range of performances. As the old saying goes, 'A chain is only as strong as its weakest link', and there will always be one component that holds back overall performance in a particular machine. Most complete systems are put together in a fairly balanced way, so for general‑purpose applications there ought to be no obvious contenders for immediate upgrading. However, the thorny subject of hard disk recording generates more telephone calls, letters, and emails to sequencer helplines (and the SOS offices!) than any other subject. The difficulty is that hard disk recording is very unlike most mainstream applications in the way that it stresses the PC. Equipment that works perfectly well in most other applications, including normal MIDI sequencing, may well stumble when called on to provide real‑time audio processing as well as simultaneously recording and playing back tens of megabytes of data from the hard disk drive.

It's a problem that's cropped up frequently in SOS over the last six months, in the ongoing discussion of the relative merits of various processing chips, and the performance of certain makes of hard disk. To shed more light on the subject, and to help those who need an upgrade, but are not sure of what, here are full details of which components affect hard disk recording. Also included is a head‑to‑head benchtest of PC processors running actual music applications, showing just what happens in real life. This will help you to find out whether you are suffering from the Bottleneck Blues.

Slipped Disks

More than any other component in the PC, the drive gets 'thrashed' when you're hard disk recording. Called upon to play back eight or more tracks of existing audio, which may well be scattered all over the surface of the disk, it must also simultaneously be able to write the audio data that you are recording. Mainstream applications normally require short bursts of data to be loaded or saved, with a lot of dawdling in between, while the user types in further information or admires the screen display, so hard drives are still widely specified using 'burst' ratings. But a quick sprinter is not as useful for hard disk recording. What you need is a marathon runner — not as fast over a 100‑yard stretch, but much more likely to get to the end of a 26‑mile course without having to stop for some muscle manipulation.

The most useful specification for our purposes is 'Average Access Time', since this specifies the total time taken to find the data required. Many manufacturers specify 'Seek Time' (a good figure to aim for is 10ms or better), but this figure only gives the time taken to move the read/write head to the correct position, and is not the whole story. To convert from Seek to Access Time, you need to add the time taken for the disk drive to rotate enough for the correct data to arrive beneath the head. This is known as Latency, and is determined by the spindle rotation speed, since the faster a drive rotates, the sooner any bit of data will arrive — on average — under the read heads of the hard drive. Latency is calculated from the rotation speed, as the time for half a rotation. Most well‑specified drives currently have a speed of 5400rpm (giving a latency of 5.56ms), and will be good enough for hard disk recording. The Quantum Fireball drives, which were the darlings of the PC magazines in 1996, used to achieve excellent overall test figures. These have, however, been replaced with the newer Fireball TM range of drives which, although they have have lower prices, compromise on performance by dropping rotation speed to 4500rpm (giving a latency of 6.67ms), and so are not so suitable for hard disk recording.

There are drives available that run at 7200rpm or even more. These should give excellent performance, but may suffer from two unfortunate side‑effects. Some have a tendency to emit a high‑pitched whine (because of the higher rotation speed), which can be annoying in the recording studio. Also, the heat generated by a drive is connected to its speed of rotation, so some drives with a very high rotation speed (such as the 10,000rpm Cheetah SCSI) may even need an additional cooling fan to ensure reliability, which is hardly welcome in the peaceful world of the recording studio.

Most people already have PCI disk controller cards, but if you have an old ISA card, then this will hold back the top performance of your drive: the PCI buss operates at 33MHz as opposed to the ISA's 8MHz, and is therefore more efficient at moving data about, but also, once started, it can continue transferring information while the processor carries on doing other things.

In the continuing debate over the merits of EIDE and SCSI drives, hard disk recording is again a factor. Although the latest EIDE drives have speeds on a par with SCSI devices, they are still likely to occupy your main PC processor more than any SCSI device will — provided it's using a bus Mastering DMA SCSI card, which has its own onboard processor to carry out the bulk of hard disk data transfer. If you want real‑time EQ and effects, using a SCSI drive could be a great way to release more of your main processing power for this purpose. As always, this rams the point home that it's pointless to look at component performance figures in isolation.

It's the same situation with 24‑speed CD‑ROM drives. Although test figures show that data can be read at up to 24 times audio CD speed, moving this amount of data from the CD‑ROM to other areas, such as your sound or graphics card, will occupy your PC's main processor for a large proportion of the time — expressed as a percentage, this is known as processor overhead. Manufacturers quote the top speeds available when the processor is occupied all the time: 100% processor overhead. But many PC magazines also test these drives in more real‑world situations, for example with an overhead of 60% (which would leave 40% of the time for the processor to actually run a real application), in which case the drive transfer rates will be significantly slower: test figures for a single component never tell the whole story.

The final factor for hard drives is thermal recalibration. All drives expand and contract as the temperature varies, so they have to periodically nip off to check that the data is still being dealt with reliably: most spend a couple of seconds every 20 minutes or so carrying out this thermal calibration. It's not usually noticeable — but, during a marathon, the last thing you want is the PC telling you that one of your limbs is going to be tied up for a few seconds doing something else. Most people carry on regardless and, on the rare occasion that their sequencer throws a wobbly, will probably not even realise what has caused the hold‑up. The chances of a peak amount of audio data being required at the same time as a thermal calibration is probably fairly low, and some drives do wait until they're idle before they do one.

However, if you want a bomb‑proof machine (and in recording studios it is commercial suicide to leave this sort of thing to chance), you will require a drive that's optimised for audio‑visual use. An AV drive might be rather more expensive than a non‑AV device, but its thermal recalibration will be specially designed to use a minimal amount of time, and this will probably be spread out, rather than all at once, so that the drive never 'stutters'. AV drives also tend to be of the SCSI variety so, although the drives are more expensive than EIDE types, the difference between AV and non‑AV is often only about £30. Micropolis is the name that crops up more often than not in this area.

Chips With Everything

For general‑purpose recording of MIDI music, even the now‑humble Pentium 100MHz processor will be more than sufficient, since MIDI does not take a large overhead. For example, I tried playing back MIDI files containing between eight and 16 fairly densely packed tracks, and with a Pentium 100MHz processor the total overhead was between 10 and 20%, using either Cubase Score or Logic Audio — no problems here! Once you introduce an audio component, though, the situation changes dramatically. If you're recording and playing back multiple audio channels, the speed of your hard disk drive becomes a far more important factor. If you look at the results of Cubase Audio's test in the 'Processors Compared' box, you can see that, whatever the processor, the bulk of the total time taken is spent reading from the hard drive.

Most manufacturers nowadays recommend a minimum of a Pentium 133 — or, more sensibly, a 166MHz — for any PC hard disk recording system. However, compatibility can be a problem. For mainstream applications, the three main processor manufacturers, Intel, Cyrix and AMD, each have a range of processors designed to have similar performances to their competitors at particular clock speeds. However, Cyrix, whose P166+ benchtests around 6% faster than the equivalent Pentium 166MHz in a range of mainstream applications (confirmed by my tests), has one area in which it finds it hard to compete — floating‑point arithmetic. Every computer processor contains an internal maths co‑processor or FPU (Floating Point Unit) to work out the results of complex mathematical calculations. Music software may use this sort of calculation a lot to ensure superior audio performance, but in mainstream applications it's rarely important, and people happily use the Cyrix P166+ instead of a straight Pentium 166, to give a slight performance boost but at about half the chip price.

When a set of audio waveforms is mixed together to produce the final stereo pair that will emerge in real time from the output of a soundcard, it seems, from my tests, that little floating point is again used, and so a Cyrix chip can compete with even an MMX in some cases. However, as soon as any EQ or effects are needed, a lot of floating‑point calculations will get carried out, and the performance penalty if you're using a Cyrix processor will be much more significant.

Your choice of processor may be dictated by other hardware requirements, as some manufacturers (notably Creative Labs with their AWE64 soundcard) have decided that, rather than put up with the slower floating‑point performance of the Cyrix chip, they will automatically disable the WaveSynth and WaveGuide features if a Cyrix processor is detected. This does rather take the decision out of your hands, but Creative Labs' stance does make it easier to make a processor choice if you must have an AWE64 card.

Hard disk recording is very unlike most mainstream applications in the way that it stresses the PC.

However, with the advent of Cakewalk Pro Audio 6.0, and the soon‑to‑be‑released PC version of Cubase VST, sequencers are increasingly being called upon to achieve real‑time processing of EQ and effects, in addition to MIDI and audio recording and playback. This uses more and more floating‑point maths, putting more and more strain on the processor. The result is that an ideal system for this sort of application will need a more powerful processor to avoid a bottleneck — and the standard benchtest results published in mainstream PC magazines can be misleading to musicians who use their PCs for hard disk recording with real‑time effects.

As far as I have been able to ascertain, there have never been any tests published to show how significant any performance penalty might be when actually running the latest music applications. So I persuaded my local dealer (Solutions of Cheltenham) to lend me a real Pentium 166MHz processor, and a Pentium 166MMX, to compare with my Cyrix P166+. The results of my investigation can be seen in the 'Processors Compared' box. However, do remember that, although you can isolate the variation due to this one component by simply swapping the processor and leaving the rest of the system alone, this variation is still no more than a guide to how much the same change will affect a different system.

Testing One Two Three

All these processors have a typical overhead when playing back most MIDI files of only 10 to 20%. Of course, if your machine is already struggling to manage eight tracks of audio, this additional MIDI overhead may tip the balance, and cause glitches. If you look at the data in the 'Processors Compared' box, Cubase Score 3.05 running on my machine with an Intel Pentium 100MHz takes 125ms per stereo track, so running eight tracks would take 8x125=1000ms, or exactly 1 second, to process each second of data. Attempting to run a sequencer as well is impossible, and in fact the Cubase performance test did show a maximum of seven possible stereo tracks. Any faster processor, including the Cyrix P166+, will be quite happy to deal with eight simultaneous stereo audio tracks. Interestingly, in general MIDI+Audio recording applications, the Cyrix P166+ outperforms the standard Pentium P166, even matching the latest P166MMX. This in itself is not a reason to buy the Cyrix chip, but it shows how efficient the Cyrix is for mainstream applications that need little floating‑point calculation. The Cubase test shows that, whichever processor is used, it's the speed of the hard disk that largely determines how many simultaneous tracks will be achieved. My Quantum TM drive is slow, and I would see significant improvements in overall performance if I replaced it with a drive that had a rotation speed of 5400rpm or more.

Wavelab 1.6, with a Cyrix P166+ in the system, will manage three chorus plug‑ins (10+24+24+24=82%), but will glitch with four in circuit. You can work out for yourself what typical combinations of plug‑ins this system will manage. In particular, note that the WaveLab reverb takes a comparatively low overhead, indicating that it's fairly basic compared with TrueVerb in the Waves Native Power Pack, but of course it does come bundled free with WaveLab, and works well in context on single tracks. With Sound Forge, normal operations such as compression and gating will work fine with a Cyrix P166+, but any heavy‑duty Paragraphic EQ or TrueVerb falls over in real‑time preview, although of course it can still be used in non real‑time.

For many people, possibly the most surprising results will be those for the MMX processor — in most of the music applications I tried, there was absolutely no difference between a straight P166 and the MMX equivalent. Only when looking at a sequencer in action (see Cubase Audio's performance figures) did I find the overall 10% or so improvement claimed for mainstream applications. You would be mad to buy a new system without an MMX version installed, but if you intend to just upgrade your processor, don't bother to move sideways to MMX — go for a higher clock speed, as this will result in a much better performance. Overall, the results are perhaps not that surprising. With my machine running standard MIDI sequencers, or even up to eight channels of stereo audio, the difference between these processors is dwarfed by the basic performance of the hard drive, which is doing most of the work, and is the main bottleneck.

However, the situation changes dramatically as soon as real‑time processing enters the scene — typically, for amplitude‑based manipulation (such as compression, gating or the L1 Ultramaximizer), the true Pentium chip outperforms the equivalently clocked Cyrix chip by between two and three times. When a lot of floating‑point maths is needed (for instance, by the Paragraphic EQ), this performance boost by using a true Pentium leaps to 360%, and for real‑time EQ applications you could compare the performance of the Cyrix P166+ chip with a Pentium 66MHz!

I was hoping to include processor tests using the only currently available sequencer that allows real‑time EQ and effects (Cakewalk Pro Audio 6.0) but although this worked extremely well, it's a 'multi‑threaded 32‑bit' application: it typically occupies no more processor time than similar applications, but the processor overhead utilities permanently read 100%, making it impossible to get meaningful figures.

Other Links In The Chain

The main potential bottlenecks for PC hard disk recording do tend to be the processor speed and the hard disk drive, but the rest of the components will have some impact too. If individual differences are small — the overall performance spread for motherboard designs, for instance, is only a few per cent from slowest to fastest — they still have a cumulative effect, so the company that markets the fastest PC with a particular spec will be the one that chooses the fastest component in each category. This will obviously affect price; if you buy the cheapest PC available, you could find it up to 10% slower than the best. Finding out which components tend to achieve the highest scores in isolation is a start, but building a balanced system requires more than one pedigree component.

The Matrox Mystique and Millennium graphics cards, for instance, have become almost standard in well‑performing PCs, but at very cost‑effective prices. The last thing you want holding up overall performance is a screen update, and both hardware design and driver efficiency will affect this. In many cases, using a display with 256 colours rather than 64,000 will reduce your graphic overhead significantly, and using a utility such as QuickRes (see May's PC Notes) means that if you need alternative colour depths or screen resolutions for other non‑sequencing applications, they are only a couple of mouse clicks away.

No‑one should be using 8Mb of RAM nowadays, unless they like their PC to limp. 16Mb is adequate for many purposes, although 24Mb is better, and, with memory currently costing about £50 per 16Mb, a total of 32Mb is slowly but surely becoming the new standard. Performance increases are claimed to be between 5 and 15% when you move from 16 to 32Mb, but, again, this is in mainstream applications. EDO (Extended Data Output) RAM increases the performance of a PC by allowing the next read to start before the previous one has finished (writes happen at exactly the same speed as normal RAM); this overlap cuts the cycle time by about 20%. The motherboard must support it, but the chances are that, if it does, EDO RAM will already be fitted (check your BIOS readout). The pipeline cache is normally claimed to give typical performance increases of around 3% when expanded from 256Kb to the maximum 512Kb. Since this normally only costs about £15, there's no reason not to upgrade, as long as your motherboard has an expansion socket to take the Cache On A STick (COAST) module. It's best to buy this from the supplier of your motherboard or complete PC, since various flavours exist, depending on the motherboard model.

A Question Of Balance

Whatever the spec of your PC, there will always be one component that's the weakest link in the chain. Judging by readers' letters, often the only difference between a budget PC hard disk system and a professional one is that the soundcard is borrowed! Regarding MMX, although the potential performance boost with some applications is claimed to be as high as 60%, in reality, six months after its launch, there are still few applications that take full advantage of the possibilities. Games and graphic packages stand to gain the most but, for music, none of the major manufacturers seems intent on using MMX. This is partly because 24‑ and 32‑bit audio manipulation is needed for high‑end music applications, with a lot of floating‑point maths. The extra features provided by MMX are activated internally as an alternative to the normal floating‑point unit, and so there will be some performance penalty if you're constantly switching back and forth between the FPU and MMX features. Ultimately, the main improvement if you are buying an MMX processor is the 10% or so basic speed improvement, caused mainly by its other features, such as a bigger internal memory cache. The tests here show that, in real terms, MIDI+Audio sequencers will show this sort of improvement, but floating‑point work will show no noticeable changes.

No doubt the subject of compatibility will come back to haunt us when AMD's new K6 processor appears in large quantities. This has a typical performance that beats the Pentium MMX processor at its own game and, since the MMX code is identical to kosher Intel devices (having been directly licensed from Intel themselves), there ought not to be any compatibility problems. We'll wait and see. Cyrix have a new range of M2 processors due to be launched shortly, but their version of MMX is re‑engineered, to provide the same functionality but with different code. It sounds as if there could still be troubled waters ahead for Cyrix‑owning musicians.

The moral is clear — if you intend to run hard‑disk audio then you can get away with using the Cyrix P166+, and save yourself about £100. If you anticipate doing any EQ or real‑time processing (Sound Forge, Wavelab, Cakewalk Audio v6, Cubase VST) then a real Pentium is a must. I'd better start saving up!

Thanks to Solutions of Cheltenham, and in particular to Dave Bruce, for their help.