Introduction
A quick look at Intel’s latest road map and AMD’s promises at CeBIT 97 leave no doubts that the times of 66 MHz system clock frequency, introduced with the birth of the Pentium processor, are going to be over soon. A memory throughput of only 528 MB/s puts the breaks on the Advanced Graphics Bus AGP, and as long as the processor accesses memory at a width of 64 bits instead of 128 bits the only feasible solution is an increase in the system clock frequency. The tinkers and overclockers among the readers of my website know since November 1996 what kind of a substantial performance gain can be had from an increase of the system clock from 66 to 75 or even 83 MHz; and only now first systems with 100 MHz system bus are slowly evolving. Actually, a jump in the system clock frequency from 66 to 100 MHz doesn’t seem that impressive, it being equivalent to a mere 50% increase, as opposed to the 100% increase from 33 to 66 MHz at the time of the switch over between the 486 and the Pentium processor. While the type of RAM in use at that time could easily be adapted to this doubling of the clock frequency, the increase to 100 MHz approaches the technical limits of currently used RAM, and a doubling to 133 MHz will probably have to wait for some time. In all likelihood a system clock frequency of 100 MHz will mean the death of FPM (Fast Page Mode) and EDO (Enhanced Data Out) RAM, which already has trouble with a system clock frequency slightly in excess of 66 MHz. Until direct RDRAM and SLDRAM, both still under development will become available and allow for extremely high clock frequencies, systems with 100 MHz system frequency will primarily be based on SDRAM. However, the quality requirements of this type of RAM will have to be raised as well before this happens. Right now there is a pretty big chaos in the SDRAM market and it is not all rare occurrence that a given type of SDRAM works only in few motherboards, or even none at all. This has to change, because systems running at a 100 MHz need cleaner timing than the 66 MHz systems in use today.
It is already well known, that 100 MHz system bus will significantly increase the overall performance of Socket 7 systems, which is merely due to the fact that the speed of the Level 2 cache is directly coupled to the system bus clock in these systems. With Intel’s Pentium II CPUs the story is a lot different, since the Pentium II has got a L2 cache that runs at half the CPU core clock and hence completely independently of the system clock. An increase in system bus clock does therefore only improve the speed of memory reads and writes, normal office applications as well as most games, that mainly run within the 512 kB of the Pentium II Level 2 cache can only benefit from a pretty limited performance gain. This fact is the reason why I was disapproving the overclocking of Pentium II systems to bus speeds of 75 MHz or even more, unless there was a gain in CPU clock frequency as well.
Now before you start wondering what this increase in system clock from 66 to 100 MHz for Pentium II systems, offered by Intel’s upcoming 440BX chipset, is supposed to be good for, let me remind you of what I said about the limitations of AGP. Although AGP is offering a theoretical data bandwidth of 528 MB/s in 2x mode, this value can impossibly be reached in systems with 66 MHz system bus, because the whole memory bandwidth of current systems at 66 MHz system bus is also exactly 528 MB/s. AGP will only get what’s left of this bandwidth after the CPU and e.g. the DMA controller were taking their share. Increasing memory bandwidth to around 800 Mb/s by raising the memory clock from 66 to 100 MHz will give AGP the ability to really take advantage of the 2x mode. This will become noticeable when finally there will be some applications that are heavily using this feature, like e.g. games with a lot of very large textures. Of course there is still some performance increase in normal mainstream applications, because the data needed by an application doesn’t always come from the L2 cache and the L2 cache itself can be ‘filled’ with new data at a higher speed as well. However this will hardly reach any percentages in the two digit range. As applications use more data and the L2 cache will get overridden more often in the future, the increase in memory bandwidth is inevitable. However this new technology won’t provide a vast increase in performance right now, just as in case of AGP, where the performance gain over PCI graphic cards is also pretty small currently. Let me again remind you that this story is different for Socket 7 systems, where the gain in L2 cache speed is responsible for 90% of the performance increase. The a lot more advanced technology of the Pentium II with its so called ‘dual independent bus’, pointing out that memory and L2 cache are de-coupled from each other, does not benefit from the system bus speed increase as much as Socket 7 systems.
A Few Words about the BX Chipset
The main difference of the 440BX chipset over the well known 440LX chipset is its ability to run at 100 MHz system or ‘front’ bus. BX also comes with a slightly different PIIX, the PIIX4e, equivalent to the PIIX4 and only adding 100 MHz support. The BX chipset is supposed to be teamed up with PIIX6 once its finsihed (probably not even this year anymore), which will offer support for IEEE 1394 or ‘FireWire’, a new and pretty revolutional bus system. A lot of people asked me about the clock speeds PCI and AGP are running at in BX based systems. To answer with a little latin term, ‘per definitionem’ the PCI bus is supposed to run at 33 MHz as well as the AGP has to run at 66 MHz, this is of course the case for BX systems also. Hence there is no need for new AGP or PCI devices, they are running just the same as before. This is another reason why the benchmark results of 100 MHz ‘front bus’ (the Intel term for ‘system bus’) systems are not even as high as in LX systems that are overclocked to 75 MHz front bus or more. In these overclocked systems the PCI bus as well as the AGP are running above their specifications, which is often enough reason for trouble, but also offering a better data bandwidth of the PCI bus or the AGP as long as it works. Neither a higher data bandwidth nor the trouble will be found in BX based systems, not even at 100 MHz front bus. You may also wonder if BX based motherboards will be able to run at 75, 83, 92, … MHz front bus as well. The answer is a clear ‘NO’ in the first place, although it will depend on the motherboard manufacturer to implement some tricks. The BX chipset is designed to read out the front bus speed from the Pentium II CPU in the system, finding out if this CPU is designed for 66 or 100 MHz front bus. Thus the systems are supposed to only either run at 66 or at 100 MHz front bus, depending on the CPU thats used. This feature is the original spec of Intel and it depends on the motherboard manufacturer if we will have the chance of running a 66 MHz front bus CPU at 100 MHz by setting a jumper, or if we will have to face the fact that a 333 MHz Deschutes won’t even run at 350 MHz. The board I tested was equipped with a special wire to fool the BX chipset, but that will most likely be removed in the final version. Looks as if die hard overclockers will need their good old soldering iron in the future.
The BX chipset as well as 350/100 and 400/100 MHz Deschutes CPUs are supposed to be released April 15, 1998.
The Benchmark Results
The results are not supposed to give you an absolute measure of how well systems with BX chipset will perform in comparison to systems with previous chipset models, because the tested board is pre-release and therefore the achieved results can be a lot different from what we will see when 440BX systems are final and hit the market. This test is only meant to give you some idea of how much performance gain can be expected from the raise in system clock from 66 to 100 MHz. Please also note that unfortunately I could not run the tests with more than 32 MB RAM at 100 MHz system clock, which minimizes the influence of the increase in memory bandwidth on the overall system performance even more. These results are only supposed to put an end to the old discussion if 100 MHz system bus will increase overall performance of Pentium II systems by a significant amount or not. You will see that my prediction was correct, the increase in overall performance in common applications is not much, but this is simply a mathematical equation rather than some mystical occurrence, as a lot of other people that publish about PC hardware wanted to make us believe. Intel is taking another step ahead into the next century, the impact of higher memory bandwidth may not be visible right now, but you can be sure that even in the close future computer systems will require at least 100 MHz memory clock, be sure that 200 MHz or even more will be realistic within the next 5 years.
Test System:
- BX board
- Deschutes CPU
- 32 MB DIMM Samsung
- nVidia RIVA 128 AGP reference card, driver 4.10.01.0230
- Quantum3D Obsidian 100SB 4440
- IBM DAQA 33240
- Windows 95 OSR 2.1
Results:
The test system with only 32 MB RAM is showing only a small performance gain in office applications by the 100 MHz system bus. The performance gain should be higher when using 64 MB RAM.
Quake II benefits by about 6%.
The performance gain of in GLQuake is a little less with about 5%.
Incoming at high resolution was supposed to show a difference due to the usage of large textures. However, the performance increase is a mere 3.5% and the frame rate doesn’t seem to be depending on CPU performance at all, it’s only the system bus that makes a difference.
I hoped that this test would show a larger performance increase when using 100 MHz system bus clock than the above tests. This is a little bit surprising though, since the RIVA only uses AGP 1x mode. Nevertheless is the impact of higher memory bandwidth in this test almost 8% and hence more than in any other test I performed.
3D Studio Max is heavily FPU dependent, but the calculation is purely running within the L2 cache. Hence there is practically no difference between 66 and 100 MHz system bus clock, the rendertime scales pretty much linearily.
You will certainly realize that it will take different applications to see a performance benefit of more than 5-8 % from switching from 66 to 100 MHz system bus clock in a Pentium II system. These will be applications that use large amounts of memory either directly or for caching. Data base applications as well as server environments are typical examples. There will not be that much of a difference for the normal home user though, at least not for now.