Dealer Focus: Processors

Intel was absolute master of the market when the only processors available for the PC were the 8086 and 286. But with the arrival of the 386 the situation changed dramatically and clone processors began to proliferate. The battle is still in progress and Intel continues to dominate, with its Pentium II looming on the horizon. But how much longer can it hold the high ground?

The idea behind the x86 processor has always been to produce a CPU which is compatible with earlier ones, but not necessarily faster. For example, Intel went from a 16-bit bus in the 286 to a 32-bit bus in the 386 and introduced level one cache inside the 486. With the Pentium, in addition to increasing the amount of level one cache ? rising from 8Kb in the 486 to 16Kb in the Pentium ? the company introduced elements of technology taken from Risc processors.

The Pentium has three lines of execution or pipelines: two for the integer unit (main and secondary) and an other for the floating point. In this way the microprocessor can carry out two instructions at the same time (in parallel). The main pipeline can execute any kind of instruction, while the secondary one has restrictions.

When two instructions arrive, the first one is put into the primary pipeline and the second one in the secondary pipeline. If parallel execution is impossible, the second instruction is put into the primary pipeline once it has carried out the first instruction. When instructions are executed in both pipelines, the result appears as if it had been carried out sequentially (first the first one and then the second) although in reality this is not the case. In this way, software runs faster, especially if it has been compiled for the Pentium.

Pentium compatibles

Cyrix has perfected execution of the 16-bit and 32-bit code, using up to 20 known general purpose search techniques, with two super pipelines of integers. As in the Pentium there is a system of prediction and speculative execution of instructions.

The company?s next chip, the M2, is based on the 6x86. It will be adapted for 32-bit software, in much the same way as the Pentium Pro, although it doesn?t suffer the same problems when working with 16-bit code, and even contains a set of MMX instructions.

For its part, AMD has developed its AMD K5 ? based on Nexgen?s Nx686, which it acquired in the first half of 1996 ? with an internal cache of 24Kb.

Speed freaks

All the chips give rise to some potential confusion about clock speeds. The Cyrix P150+ really functions at 120MHz, the P166+ at 133MHz and the P200+ at 150MHz. Cyrix maintains that its chips run faster than the Intel equivalents and it needs to position them without referring to their actual speeds. For example, according to Cyrix, its P200+ performs better than a Pentium 200, but if they were compared on real clock speeds alone, the assumption would be that the Intel chip is the faster one.

The 6x86 gives off much more heat than the Pentium and the AMD K5 because of its internal design. Cyrix points out that the 6x86 machines have to use a more powerful cooling mechanism than Pentium machines.

Put it to the test

The tests are designed to give a reliable indication of each chip in different circumstances. The first, 3D Bench, is a simple application which measures the performance of 3D system graphics. The second, the compilation of the Linux kernel, is one of the most demanding and complete tests available.

The third test used is the index of frames per second of Quake, although this test is not particularly indicative of the real performance of the machines.

To measure the floating point capabilities of the coprocessor, we performed a ray tracing test using the program Pov-Ray.

In the fifth and last test, we measured the Mips, MFlops and rates of data transfer to the video card and to the main system memory. This information was used to build an index for each processor in the following way:

• We counted the Mips as 40 per cent of the final mark. In reality we should have made the percentage nearer to 60 per cent rather than 40 per cent, but we decided to spread the pure processing mark between the rest of the variables to balance the comparison more evenly.
• The floating point unit or mathematical coprocessor MFlops were given 20 per cent of the final mark. Although most common applications (videos, multimedia, Mpeg or AVI, image treatment, text processing, communications, Net servers etc) do not require much calculating power, there are tasks such as rendering and use of 3D images where a good coprocessor is almost more important than the CPU.
• Another 20 per cent was attributed to Ram accessing, a basic process on any PC and one which slows down the system to a greater or lesser extent, depending on motherboard design and cache.
• The last 20 per cent corresponds to video speed. In reality, it is not that important a factor in measuring the power of a processor, but rapid screen updates always make a machine appear faster.

The base boards used for the tests were an Asustek with the Intel 430HX chipset and another OEM model with the Via Apollo chipset. Both had 48Mb of EDO Ram and a STP Velocity 3D graphic controller with 8Mb of VRam.

Get a result

None of the processors presented any incompatibility problems with Windows NT 4, OS/2 Warp 4 and Linux Redhat 4 during the tests. In general it can be said that they all performed excellently and that in no case could it be said that the equipment was slow and it must be remembered that the worst of them had twice the power of a 486DX2/66.

The 6x86 bug

During the test phase of Windows NT 4, it was discovered that some Cyrix 6x86 processors provoked random system crashes. Microsoft investigators went to work and discovered that these crashes were caused by a bug in the design of the cache writeback of the 6x86 and that it could be cured by deactivating the internal cache from the user in the defective series of the microprocessor.

Deactivation of the cache, however, results in a fall in performance of about 40 per cent. Cyrix has solved this problem, and versions of the chip from 2.7 onwards function perfectly with NT4.

Motherboard superior

We used three motherboards, each with different chipsets: an Asustek, an OEM type and a board from DF1.

The Asustek P55-T2P4 with Intel 430HX chipset supports an operating frequency of up to 75MHz, which is vital to make full use of the Cyrix 6x86 P200+. Out of the three boards, the Asustek turned in an excellent performance every time.

The OEM type incorporated the Apollo chipset, made by Via. It performed excellently in our test, getting close to, and sometimes beating, the performance of the Asustek.

The G587VPS Pro from DF1 is based on the VLSI Lynx chipset. However, it arrived after we had finished most of the tests, so to some extent we have had to rely on initial impressions.

Asustek?s board performs five to 10 per cent better than other boards with the 430HX chipset, even those from manufacturers as sound as Gigabyte and Soyo. The Asustek is thus not representative of boards using the HX chipset.

As far as the tests comparing the two chipsets are concerned, the graphs on page 47 show how both boards provide very similar performances with the Intel and the AMD processors. But with the Cyrix 6x86, the performance of the board with the Via Apollo chipset is inferior to that of the Asustek with the HX. This is strange, since the manufacturer claims it is especially suited to such microprocessors ? which seems to be contradicted by the test results.

Test results

Pentium MMX

The Pentium MMX is the new generation of processor from Intel. It differs from previous ones in that it incorporates a new set of instructions to speed up multimedia operations. But existing multimedia applications do not use the MMX extension and therefore do not gain any advantage from these, although it is certain that all new multimedia applications will use MMX in the near future.

The MMX also offers more power than normal Pentiums, since Intel has doubled the internal cache of the chip to 32Kb. This was amply demonstrated in the tests on the 166MHz Pentium and 166MHz MMX, in which the MMX outshone the plain Pentium.

As far as price and performance is concerned it makes sense to buy the 166MHz Pentium MMX rather than a 166MHz Pentium.

AMD K5

The results of the tests might give rise to the impression that AMD K5 chips are slow and are not worth considering. But personal experiences with the chips are much more positive than the figures indicate.

The AMD-PR133 is the processor which we liked the most, after having used it in our system in OS/2, NT and in Unix environments. It performs differently under Windows 95. The processor is smooth and its 24Kb of level one internal cache is an important factor in its performance.

In fact, as far as the Pentium 120 is concerned, the machine is faster with the AMD K5-PR133 and yet, according to the results of our tests, it is not as good. Apart from its weak spot of floating point calculations, the chip can generally be recommended in every respect. It is a pity that there is no K5-PR166 yet ? if it ran as smoothly as the PR133 in 32-bit systems, it would be an option for the mid-range.

Intel Pentium

As far as the Intel Pentium is concerned, it turns in excellent floating point performance, thanks to its coprocessor which is far superior to those of its competitors.

Apart from the coprocessor, in many other applications (internet servers, local network, text processor, communications and so on) the performance in integers (CPU, pure) is also very good. The chip is the recommended option if price is not a consideration.

Cyrix 6x86

In the case of the Cyrix 6x86, we discovered processors which go like a rocket with certain applications and are very competitive on price. They are also a good alternative to bear in mind as general purpose processors.

Like the AMD K5, the Cyrix coprocessor leaves something to be desired with respect to the Pentium. According to tests, the floating point unit of the 6x86 obtains a performance of about 60 per cent of that of a Pentium, although it is 10 to 15 per cent better than the K5.

Nevertheless, the power is generally more than enough to offer a very respectable performance with floating point calculations, making the 6x86 a good choice for cheaply updating a computer. The 133MHz version, marketed as the 6x86 P166+ because its performance is comparable with that of a 166MHz Pentium, comes out on top as far as performance for price is concerned. But there are still some reservations.

First, not all Pentium boards support the 6x86. In our tests it would not work with a motherboard that used the 430FX chipset, whereas the Pentium and AMD K5 worked perfectly. We presume this is because it needs a special Bios on the motherboard to initialise it and enable the equipment to start operating normally. This does not necessarily mean the 6x86 will not function in any old motherboard, but it does mean that it will not automatically work in them all. In general, if the board has been manufactured during the second half of last year, it will almost certainly incorporate support for the processor.

Another slight hindrance is the NT 4 bug. In reality, it is not a bug in Windows NT, but in the Cyrix/IBM processor, since, under certain occasional special conditions, the writeback cache of the 6x86 processor makes an error. In order to avoid it, Microsoft has introduced code which detects what version the Cyrix/IBM processor is. If it is earlier than version 2.7 then NT 4 deactivates the 6x86 internal cache.

Another possible inconvenience is heat dispersal. The design of the 6x86 processor means it emits far more heat than a Pentium or an AMD K5, and Cyrix acknowledges this. Heat generation is more evident in the 3.5V units than the 2.8V models. There have been cases of distributors selling the 6x86 with only a regular fan instead of a special one, so it is worth checking.

In addition, the 6x86 P200+, which runs at 150MHz, has a particular disadvantage because the frequency at which it and the motherboard work are fixed and the multiplication factor cannot be changed like it can with a Pentium. This also occurs with the AMD, but with these processors it does not pose a problem. The 6x86 multiplier is limited to duplicating the frequency of the motherboard, which restricts it to a factor of two-and-a-half or three. This means that for the 6x86 to function at 150MHz, the board is required to work at 75MHz. And there is the problem: not all Pentium boards function at 75MHz ? they work at 50, 60 or 66MHz. So on a board that does not support 75MHz, the chip cannot work faster than 133MHz. Anyone buying a P200+ therefore needs to check whether the motherboard can support such a speed.

None the less, for excellent performance at a modest price, the 6x86 P200+ is an excellent choice on a motherboard with 75MHz support. Of course the performance is nothing like that of the Pentium 200, but on occasions it is not far off. In our opinion, as far as price and performance is concerned, it is one of the best choices.

Although a 200MHz Pentium is more powerful, it has been proved that some machines running Windows 95 or NT using a 6x86 P200+ are faster than a Pentium-based machine using the same operating systems. Strange, but true.

Conclusions

The most attractive options are the AMD-K5-Pr120 for the lower range, and the 6x86-P200+ and the Pentium 150MHz for the medium range. In the upper range, the laurel wreath must go to the Intel Pentium MMX, although the large price difference between the 166MHz and 200MHz chips does not correspond to the difference in power between the two. For that reason we would opt for the slower of the two.