Speeding Up Your PC For Music Applications

Whatever the processor speed, efficiency of reading and writing hard disk audio files is largely dependent on the drive itself. With all real‑time effects disabled, the main processor involvement is in mixing down multiple tracks to a final stereo soundcard output.

If you're still under the impression that the Celeron is a type of vegetable, now's the time to find out more. Martin Walker plugs in some new processors.

The world of the PC changes so rapidly that since I last covered some of the hardware choices (as part of the 'Going a Bundle' feature in the October '98 issue of SOS) there's a whole new range of options available. In fact, some items that I recommended for budget systems (such as the Pentium MMX processor) are no longer available at all, having been sacrificed on the altar of progress. Of course, at the same time software developers are providing us with more and more enticing options to consume more of our processor power, fill more of our hard drives, and deplete our bank balances.

Deciding which is the best system upgrade is always tricky, especially when there is always a faster component around the corner. It is impossible to stay ahead of the game for long, and many people have already resigned themselves to the fact that whatever they buy will probably be declared obsolete within six months. However, if you're looking for faster PC performance, the time has come to take stock, investigate some of the new options, and take a look at what's coming in 1999.

Thanks For The Memory

Floating‑point performance is largely what determines how fast processors will run real‑time plug‑ins. Here I have compared all results with the Pentium II 450MHz, so you can see how much performance differs from other processors. Notice that the new Celeron range with L2 cache memory performs almost identically to a Pentium II processor with the same clock speed.

Apart from obvious factors such as hard disk performance and processor speed, the amount of available RAM memory you have can also affect the speed of your system. Windows loves to use virtual memory (the swap file on your hard drive) so that you can run far more applications than you could ever fit into physical RAM. As soon as your RAM has been filled up, some of the data in it will need to be put into the swap file when you launch another application (most of each application can be temporarily swapped to the hard disk swap file when it is not being actively used). Of course, this will slow down the launch of the new application, so most people notice a dramatic improvement in launch times after installing more RAM, if their system was previously rather starved of it.

Another classic sign of insufficient RAM is when you can run several applications quite happily, but it takes some seconds (and lots of hard‑drive activity) to switch between them. This is a sure sign that a lot of virtual memory is being used, and that plenty of data is being swapped back from your hard drive. Remember that this is also a two‑way process, since not only is the data already paged out to the hard‑disk swap file being pulled back into RAM, but Windows also has to send other information back into the swap file to make way for it. This is a recipe for audio glitches.

Each application you run will also need a certain minimum amount of permanent real (rather than virtual) memory. This is known as Locked memory, and is the part of the application that cannot be swapped onto the hard drive. Any application which is streaming audio would like to 'lock' large portions of your RAM — particularly software‑based samplers and synths, since they need to create RAM buffers for their waveforms or samples to ensure low latency (the time between pressing a key and hearing the sound). The more simultaneous notes you want to achieve, the more memory buffers are needed. If, due to a lack of RAM, new synth waveforms have to be re‑calculated every time you press a note, or new samples have to be re‑loaded from hard disk on every occasion they are required, glitches are much more likely to occur.

Bigger And Better?

The Level 2 cache affects floating‑point performance in a big way. Here you can see the result of disabling (green) and enabling (blue) the L2 cache of my Pentium II 300MHz processor. The biggest changes are for the reverb plug‑ins, which show up to a 34 percent improvement with the cache enabled.

Over the last few months I've reviewed several applications of the type mentioned above, which ideally need large amounts of RAM. The first was Gigasampler (see SOS December 1998), with a recommended 128Mb RAM requirement to reduce latency. Although you could get by with 32Mb, total number of samples, and therefore maximum polyphony, would be restricted. The second was Creamware's Pulsar DSP environment (see the last issue of SOS), and this time the recommendation of 128Mb of RAM seemed even more important, since running with only 64Mb left the environment slow and unresponsive, despite the fact that the audio part of the code was running inside the DSP chips.

If plenty of RAM is available, applications such as these should manage much lower latency values and higher polyphony without glitching. Seer Systems' Reality software synth even provides a 'Lock Memory' tick box, which can either be set to Always (for best performance), or On Demand (uses less RAM). In general, it's wise to heed the recommendations of the software developer, and to listen for sudden excessive hard drive activity if you're not sure whether you have enough RAM installed. There are several utilities (including the free Microsoft System Monitor) that let you see how big your swap file is, but it is always worth checking in the Preferences section of your software to see if there are any RAM options. Wavelab, for example, has two caching methods. The one for small to medium files (up to about 20Mb) uses the standard Windows system, but if you select the one for huge files, Wavelab will then manage RAM itself. This can make a big difference — if you open a 10‑minute stereo file (about 100Mb) using the 'Small' setting, Windows will add 100Mb to the swap file. The best option is to let Wavelab choose for each file. Taking a little time to study the buffering options in your own software will help you to get the most from available RAM.

Processors

Here's the one that most PC musicians have been waiting for — a comparison between the Pentium II 400MHz and the new Celeron 400MHz processors, both running DirectShow real‑time audio plug‑ins. For the simpler plug‑ins there is no difference, but the bigger cache of the Pentium II 400 does benefit effects that need to calculate algorithms over a longer period, such as reverbs.

Choosing the best combination of processor manufacturer, model and clock speed can greatly influence the final price of a new PC, particularly at the top end, where the 450MHz Pentium II is still priced at about £420 at the time of writing. At the other end of the price scale there is only really one serious contender for music purposes — the AMD K6‑2 — but the new Celeron processors, operating at a similar clock speed when running music applications, seriously outclass even this.

Deciding which is the best system upgrade is always tricky, especially when there is always a faster component around the corner.

I spotted some comparisons between the 350MHz AMD K6‑2 and the Celeron 333MHz (with cache — see 'Cache It If You Can' box above) run by Harmony Central (www.harmonycentral.com). These comparisons had been made using music software — Sound Forge running non‑real‑time calculations with noise reduction, pitch‑shifting, and acoustic modelling. With the two PC systems set up to be as similar as possible, pitch‑shifting proved to be 53 percent faster with the Celeron, while Celeron‑based noise reduction and acoustic modelling calculations were around 100 percent faster. I know that people with non‑Intel processors don't like to hear this, but despite the fact that the AMD range regularly pushes ahead in most office and some game software benchmarks, this is always at the expense of floating‑point performance. I've seen references to DirectShow and VST audio effects consisting of "nearly 100 percent floating‑point calculations". In addition, the real‑time mixdown of multiple tracks inside most MIDI + Audio sequencers uses floating‑point maths as well, so even if you avoid audio plug‑ins you're still likely to benefit from an Intel processor. If you're running a MIDI‑only sequencer, an Intel processor will make little difference, and if you're prepared to forgo real‑time operation for your audio effects you'll get by with an AMD K6‑2, but its extra 3D processor instructions will only benefit 3D graphics software, and only then if the software has been specially written to take advantage of them.

Intel Monopoly

This graph highlights processor differences by effect type. The Spectraliser needs a huge number of calculations over a short period, and therefore benefits from the double‑speed cache of the Celeron. Effects involving EQ and dynamics operate on short sections of the waveform, and there is little difference between the two. The larger cache of the Pentium II scores when calculating any effect that involves longer durations of waveform, such as delays and reverbs

Sadly, other manufacturers fare even worse, leaving Intel as the only serious choice for musicians. Until recently the Pentium MMX processor scraped in for music purposes, and the Pentium 166MHz MMX is still quoted by many music software developers as providing entry‑level performance. However, I suspect that many of them do this rather tongue‑in‑cheek, to avoid too many users grumbling about escalating software requirements. A more realistic proposition is a 200 or 233MHz model, but this is now rather academic, as Intel have ceased production of the basic Pentium MMX range.

This leaves two Intel product ranges — the Celeron and the Pentium II — both of which use the newer Slot 1 architecture (although this is changing as well, as I explain in the 'Socket To Me' box on page 184). Again, until recently it was not difficult to choose between these two processors. The first Celeron models had no onboard cache memory, resulting in fairly lacklustre performance. However, it was the speed of the cache memory that prevented many Pentium II processors being successfully overclocked (see the 'Overclocking' box below for more details). Since the Celeron had no cache, many people did manage to overclock them to provide equivalent performance to the much more expensive Pentium II range, and this made them popular with PC enthusiasts.

However, the Celeron was still generally seen by the industry as suitable only for low‑cost, entry‑level PC systems, so a new Celeron 'A' range soon appeared, this time with 128K of cache memory (compared to the 512K of the Pentium II models). As expected, performance was better, although the range still doesn't seem to have caught on in a big way with PC suppliers. This is probably because for most applications, including games, the AMD K6‑2 has the edge in performance terms. However, the Celeron still has an identical floating‑point unit (FPU) to the Pentium II, and this gives it significantly better performance than the AMD K6‑2 with the real‑time audio plug‑ins that nearly every PC musician now wants to run.

At the top end of the market, the Pentium II has seemed the obvious choice for some time, and since its floating‑point performance is significantly better than that of other manufacturers' chips, music retailers have made this the core of almost every PC music system over the last year. The simple logic was to buy as fast a Pentium II system as you could afford, and this was just what I did back in August '98, when I upgraded to a Pentium II 300MHz machine. Since then, both the 300 and 333MHz models have been dropped from the range, leaving the 350, 400, and 450MHz models, all of which support the 100MHz front‑side buss (which I discussed in the September '98 PC Notes). Of course, as is always the way, prices have been dropping, and as I write this at the beginning of February, the 350MHz Pentium II costs a very reasonable £160, although the 450MHz model is still a budget‑threatening £420.

And Then There Were Three

Nearly all music retailers are still sticking with the tried‑and‑tested Pentium II, but a different choice should be available by the time you read this, in late March. Intel have announced that their next generation of processors, formerly code‑named Katmai, will now have an official brand name of Pentium III. Their main aim with these, which should start shipping in March 1999, is apparently to widen the gap between the current low‑end Celeron and previously top‑range Pentium II processors. The price of the 450MHz Pentium III is expected to be about the same as the 450MHz Pentium II is now.

However, as soon as Pentium III chips ship, with initial clock speeds of 450 and 500MHz, Pentium II production will start to be phased out, with the current slowest 350MHz model disappearing over the next few months. A 533MHz version of the Pentium III is due for launch slightly later in 1999, and by the end of the year a 600MHz model is expected. This will also increase the system buss speed, from the existing 100MHz of the fastest Pentium II models, to 133MHz.

The Pentium III design is still largely based on that of the Pentium II, but its 512K cache will be integrated with the processor chip, as in the case of the Celeron, and will therefore run at the full processor speed for enhanced performance. In addition, it has 70 new multimedia instructions known as KNI (Katmai New Instructions, but also referred to as MMX2) claimed to improve 3D graphic, video and audio performance.

Faster processors always benefit music applications, but with MMX the new instructions could only be used instead of the floating‑point ones, since they shared the same internal registers, and this prevented them being used in most music software. This time, however, the Pentium III enhancements not only include improved floating‑point routines, but allow these to be used simultaneously with the new MMX instructions. Of course, anyone buying a Pentium III won't immediately reap the full benefits. We'll have to wait until software developers release the next updates and upgrades to see what (if any) improvements have been achieved.

In addition, there are no guarantees of compatibility for the new processor when it comes to existing systems. Apparently, Windows 98 will need a special patch to enable it to run a new mode available to the Pentium III, but this should be readily available for anyone to download. Existing Slot 1 motherboards may or may not run happily with the Pentium III range — certainly, the 133MHz system buss of the 533MHz and faster models will mean that they require a new motherboard (with a new JX chipset) and perhaps new memory, but the 450 and 500MHz models may run on your existing motherboard.

Always check before buying a new processor that it's supported by your motherboard — look in your motherboard manual, if you have one, or on the manufacturer's web site. My new PC uses the QDI Brilliant 1 board, and in January QDI posted a new flash BIOS update that supports Celeron processors up to 500MHz, and the forthcoming Pentium III series. Flashing simply writes new information over the top of the existing BIOS data in an EEPROM (Electrically Erasable Programmable ROM) chip, and most new PCs have flash capability. However, flashing BIOS is not to be undertaken lightly. First of all, only ever use the files from your motherboard web site, and make sure you pay attention to the available information — you may have an Award BIOS, but there are dozens of different motherboards using this core code from Award, and all of them have been tweaked to suit individual models of motherboard. Secondly, if anything goes wrong during the flash process, you might just end up with a totally non‑functioning PC, requiring a new BIOS chip from your system supplier. For this reason I would recommend that you never update your BIOS as a matter of course — only do it when you specifically need a new feature or bug‑fix. It's just not worth the risk.

The Big Match

I know that many musicians are particularly interested in the Celeron because of its combination of low cost and high‑speed FPU. Now, with the first new launches of 1999, the Celeron range has expanded — there are now two new models featuring clock speeds of 366 and 400MHz. The 400MHz model currently has a price roughly half that of the full Pentium II at the same clock speed. Both models can be bought in two versions: one for the familiar Slot 1 socket used by the Pentium II range, and a new Socket370 version (see the 'Socket to Me' box). This makes the choice for musicians more difficult, since you can now build a system for a similar price containing either a 400MHz Celeron or a 350MHz Pentium II. The core processors are identical, but, as mentioned earlier, the Celeron has an onboard cache of 128K running at the same speed as the processor, and the Pentium II has one of 512K that runs at half the processor speed. In addition, the entire Celeron range runs with a 66MHz FSB (front‑side buss), compared with the 100MHz buss of all remaining Pentium II processors.

There are various benchmark tests available that measure FPU performance, and since this is primarily what determines the number of audio effect plug‑ins you can run, I have compiled a graph showing how a comprehensive range of different processors compare (see page 186). This is based on 3D rendering tests that involve lots of FPU calculations, but I have correlated them as far as I can with specific audio tests, and they also bear out most of my own measurements, as we will see later. I have normalised them to the performance of the Pentium II 450MHz, so that you can easily see how much slower the others are by comparison. Of course, real‑world figures will also depend on the rest of the system, and the proportion of floating‑point calculations used by a specific plug‑in or sequencer application, but the graph will give you a rough guide to the relative performance of the various processors with music software.

You can see from the table on page 186 that the 100MHz FSB doesn't seem to have made as big an impact as Intel would like us to think — there's little difference between any Celeron and the equivalent speed of Pentium II. This is because although the Pentium II has four times as much cache memory, it runs at half the speed of the Celeron cache, which largely cancels out the effect of both the larger cache and the higher 100MHz FSB speed. My other tests confirm that the 100MHz buss gives no significant improvement by itself. You can also see that the AMD K6‑2 range, while running most day‑to‑day applications slightly faster than the equivalent Pentium II, doesn't fare so well when it comes to FPU performance — currently the fastest 400MHz model shows a similar figure to the original Celeron 300MHz model with no cache, and is beaten even by my own 300MHz Pentium II.

For a musician who is trying to get nearly all PC components working flat‑out to achieve the maximum number of simultaneous audio tracks and real‑time effects, overclocking is a recipe for disaster.

On The Level

To shed further light on the subject of Level 2 cache performance, I temporarily disabled the L2 cache on my Pentium II 300MHz processor (you can do this inside the BIOS). This effectively demotes it to the status of the older, cacheless Celeron 300. I then ran various tests on the hard drive and when running real‑time effect plug‑ins. I measured absolutely no difference in hard drive performance when running DskBench with simulated 8‑track playback, and this is no doubt because with bus Master DMA drivers (see the December '98 issue) the CPU is scarcely involved in disk transfers. However, when it comes to floating‑point performance, the L2 cache does have a more significant effect on performance. You can already see this in the table, where the Celeron 300A model with 128K cache beats the original cacheless Celeron 300 by 17.5 percent.

My tests show that the improvements given by the 512K L2 cache of the Pentium II vary quite a bit with audio plug‑ins, as you can see from the table on page 188. Operations such as EQ and dynamics all benefited by up to 14 percent (this ties in fairly well with the figures), but as expected it was the reverbs that showed the biggest improvement, with both Waves' TrueVerb and the TC Native Reverb showing just over a 30 percent drop in CPU overhead with L2 cache enabled, and an improvement of 25 percent for the Hyperprism‑DX HyperVerb. This is because reverb algorithms need to calculate multiple reflections for each point in the waveform as it decays, so there is a much better chance of the data already in the cache being required again.

Running Cubase VST with the supplied Quick Start Song (six tracks of 44.1kHz audio) and several supplied effects took 34 percent CPU overhead with the cache enabled, and this rose to 40 percent when I disabled it — a proportional increase of some 18 percent. Again, this is not due to any drop in hard‑drive performance, but to the fact that VST uses floating‑point maths when mixing down the tracks to stereo.

If you're not sure what to upgrade next, you first need to decide where your current system is most limited: if you want more tracks, buy a faster drive; if you want more effects, buy a faster processor.

The Big Test

Since there are now two Intel processors running at an identical 400MHz (at the time of writing the Celeron 400 costs about £135, and the Pentium II 400 about twice as much, at £270), I performed some comparisons with both these processors in my own PC. All other hardware components were identical. First, I ran the DskBench tests that accurately measure hard‑drive performance when playing back eight tracks of 44.1kHz audio. I've not bothered to produce a graph of the results, since they were absolutely identical with my UW SCSI hard drive whether using a Pentium II 300MHz (my own processor), the Pentium II 400MHz, or the Celeron 400MHz. As mentioned previously, this is because only a few percent of total CPU power is required during hard drive activity when using Master DMA drivers.

However, using Wavelab as a test bed to run a variety of real‑time plug‑ins provided some intriguing results (see the table on the right). Some effects gave almost identical results with the two 400MHz processors, and this bears out the overall floating‑point graph of earlier. However, when it came to reverbs, the Pentium II pulled ahead significantly, which ties in with my previous on/off cache tests, although the differences this time are far less. To make the trends more obvious, the lower table on the right shows the results for the more expensive Pentium II as a percentage difference from the Celeron. You can now easily see that for basic EQ and compression there is little to choose between the two. One interesting result is that achieved with the Spectral Designs Spectraliser, where the Celeron is nearly two percent faster. This is probably because adding harmonics to an existing waveform benefits from the faster cache of the Celeron — it's all a question of balance.

When running Cubase VST with eight tracks of audio and no effects, the processor drain varies quite a bit over time, but I measured about 29 percent for the Celeron and 27 percent for the Pentium II. This works out to a proportional drop of about seven percent in processor overhead for the Pentium II, due, I suspect, to its larger cache.

Summary

Announcements of new PC components tend to put some people into a panic, especially if they are about to buy a new machine or upgrade. I'm very happy with the current performance of my Pentium II 300MHz machine, and it will run quite enough tracks and real‑time effects for my purposes. The only occasion on which I found it sluggish was when I tried out the new Creamware Pulsar system, and this seemed largely due to a lack of RAM. I intend to upgrade to 128Mb shortly, although only really to review cutting‑edge items for SOS, because even running Cubase VST and Wavelab together is no problem with 64Mb of RAM. However, having 128Mb will enable me to spend more time attempting to run software sampling alongside Cubase VST.

If you're thinking of taking the plunge and buying a complete new system or major upgrade, remember that the number of simultaneous hard‑disk audio tracks is largely determined by the speed of the hard drive, and the number of real‑time effects by the speed of the processor. So if you're not sure what to upgrade next, you first need to decide where your current system is most limited: if you want more tracks, buy a faster drive; if you want more effects, buy a faster processor.

On the basis of my processor comparisons, I would personally buy a Celeron 400 rather than the similarly priced Pentium II 350MHz, since if you scale down my Pentium II 400MHz figures the Celeron will beat it on every count — and if you compare the Celeron 400 to the Pentium II 400, there still isn't sufficient difference between the two to warrant spending £135 more. Even with reverbs, the biggest overall difference I measured was less than two percent. Of course, once the Pentium III appears it may well be a different story, and rumours are already flying about the possible release of a 433MHz Celeron processor by March, followed by a 466MHz model in April. It's sometimes tempting to say "Stop the PC, I want to get off!"