Quiet/silent computing has become a topic of increased interest for both businesses and home computer users. In businesses with increasing worker density, cubicals for example, the combined noise of the users' workstations can reach irritating and distracting levels. Noise can also be a serious impediment to home computer use -- the "annoying buzz" is an imposition into the living space, especially as computers become more a part of life, to the point that they are often never turned off.

One of the main sources of noise in a modern computer is the fan cooling the CPU. Although substantial progress has been made in the area of quiet fans, the ideal for any noise-conscious system is passive (no forced air) cooling. This review will contrast two possible quiet systems with simple, passive cooling on their processors.

Platform A: VIA C3 933MHz

VIA Technology, apparently alone among the main CPU manufactorers, is putting both technical and marketing effort into the quiet/silent computing market. Their C3 processor, based on the purchased cyrix III core is touted as the ideal processor for low heat, low power desktop systems. The C3 is suggested to produce about 6 watts of heat in normal operation, and less than 12W in full use. Compared to 50-70 for some other platforms this is a dramatic reduction. Another major feature which helps it work towards fanless cooling is its ability to function in an emergency without a heatsink. This is validated by VIA themselves in their "beat the heat" video. In this video,they remove the heatsink from the C3 and it goes on to play Quake 3 (albeit slowly) for 24 hours without crashing.

The C3 is a socket 370 processor with a integrated heat slug. It is specified to operate at a 100 or 133MHz front side bus. The processor features two 64kB, 4 way associative L1 caches and one 64KB 4 way associative L2 cache.

The C3 core is highly optimized towards reduced heat output as this text from the C3 Ezra datasheet explains.

A few instructions dominate x86 instruction execution time.Over 90% of instruction execution time is due to a few basic x86 instructions. Most x86 instructions have no significant impact on performance.

The VIA C3 Ezra processor design optimizes performance of the most important X86 instructions while minimizing the hardware provided for the little used x86 functions (these are primarily implemented in microcode). The resulting instruction execution speed of highly used instructions is as good as or better than comparable processors. For example, the VIA C3 Ezra processor executes x86 load-ALU-store instructions in only one clock as compared to several clocks on other processors. The execution time of little-used instructions is impacted to reduce die size, but this has no effect on real application performance.

Improving clock frequency has higher leverage than improving CPI. The result of advanced computer design approaches over the last few years has been that the improvements in cycles-per-instruction (CPI) often impact MHz improvements, and certainly impact die size. Our belief is that the best way to improve total performance and keep a small low-power die is to improve MHz.

Taken from the "VIA c3 Ezra Core" Datasheet

These optimizations will have several limitations but the overall heat production of the core should be reduced as there is physically less silicon in operation. The C3 should perform well in its target applications; however, because of the microcode implementation of portions of the x86 instruction set it will perform poorly outside its focus area. Also, with the emphasis on increased clock speed (MHz) instead of doing more per clock (IPC) the C3 should not be expected to compete on a 'per clock' basis with other processors.

The Ezra core features a discrete floating point unit that operates at half of the processor clock frequency. This is unlikely to seriously affect the processors performance in its target area as business applications tend to require integer math. However, this will severly impair its ability to perform floating point intensive applications such as graphic manipulation or 3D games.

The processor is compatiable with a variety of chipsets and motherboards, but VIA suggests combining it with either their Apollo Pro or ProSavage chipsets. For this review the C3 will be installed in an Apollo Pro 133A based ECS Elitegroup P6VXM2 motherboard. This is a Micro-ATX board with a VIA694X northbridge and a VIA 686B southbridge. It features onboard AC97 sound, one AGP slot, three PCI slots and one legacy ISA slot. It has two DIMM slots supporting 512MB each for a maximum 1GB of RAM.

Platform B: Intel Pentium III 933MHz clocked at 433MHz (66fsb)

The Intel Pentium III is not a new processor. It started with the release of the Katmai series, available in speeds of 450MHz - 600MHz. Katmai's were not popular at the time. In the enthusiast market, the Athlon from AMD was gaining influence. Intel seemed to be struggling to keep up with the clock-ramping that AMD was aggressively pulling off. Intel's response to the Athlon was the Coppermine Pentium III. Featuring a die shrink to 0.18u, lower core voltage, much lower heat output, and integrated L2 cache on-die, the Coppermine was more than a match for the Athlon. Also included in this release was a switch to socket 370, much appreciated over the Slot-1 interface. The Coppermine was also the first CPU to appear in "flip-chip" format, with the silicon die covered only by a thin layer of blue ceramic.

Since then, the Coppermine has been replaced by the Tualatin core, allowing the Pentium III to run at speeds greater than 1GHz. The Tualatin is, however, unable to run in an unmodified 440BX platform. For this reason alone, a Coppermine was selected for this test.

The 440BX platform is known to many as the most stable and most efficient chipset ever designed. This is why the Abit BE6-II motherboard is used for testing the Pentium III, instead of an i815-based board. The i815 chipset has some newer features, like AGP4X and the ability to run Tualatin CPUs but, in the authors' opinions, it simply can not compare to the stability and speed of the 440BX running the same CPU and memory. As a side note, i810 and i820 were largely unpopular.

In order to stabilize the temperature of the Pentium III to allow passive cooling, the bus speed was dropped from 133MHz to 66MHz. This seriously hurts the performance of the system because the CPU now runs at 466MHz, and the memory runs at a slow PC66. The effects of this speed drop should be evident in the benchmarks. This unfortunate cooling consideration will be discussed in the next section.

Cooling the processor:

VIA reccomends using the Coolermaster DPSGN01 to cool the C3. This is a fully aluminum heatsink with slightly tapered fins. This heatsink will be required to cool the processor in a closed case. The only airflow available will be general movement caused by the fan in the powersupply. In keeping with the quiet target of this system, a Enlight EN-8146902 150W PSU will be used. This micro-ATX supply can provide it's full rated 150 Watts with a single, quiet, low-movement 80mm fan. Unfortunately, this means that the ambient air movement within the case will be very limited. These conditions are typical of a low-noise, few moving parts design.

Temperature and stability testing were all done in about 25c ambient air. Idle temperatures were taken in Windows 2000 with near to 99% processor idle measured in the task manager. Full load temperature and stability were tested using Prime 95. This program is designed to search for large prime number. In stress testing mode it 'searches' for known data and compares the results with it's predetermined answer. In this way Prime95 not only severely loads a processor, but also checks that the CPU is actually getting the correct answer.

The 933 C3 Ezra processor's reccomended maximum reccommended operating temperature limit is 70c, however it's upper limit without causing long-term damage is 110c. Although operation near this high limit will likely have adverse long term affects, this extra safety margin is very nice to have. The C3 idled in windows in the mid 30s, while in Prime95 it sat at around 60c. The C3 and VIA platform could perform Prime95 stress testing for the full tested 24 hours without issue. Although the fully loaded temperatures are higher than ideal, in normal operation the processor temperature is easily within all safe margins.

The P-III Coppermine is specified to a maximum junction temperature of 77c. There is always a temperature difference between measured temperature, and the hottest point in the core so the measured temperature should never be allowed to actually reach the maximum junction temperature. At 933MHz fully loaded P-III entered thermal runaway and it's temperature rose constantly to over 60c. At this point testing was terminated to avoid locking up the processor. In order to have the CPU hold a safe temperature while under load it was underclocked from it's normal 7x133 to 7x66MHz. At the resulting 466MHz it idled in the mid 30s and was stable fully loaded in the mid 50s. Unfortunately, decreasing the speed of the P-III in this manner has the added effect of dramatically reducing memory bandwith due to a 50% reduction in memory speed. Although the 466MHz P-III's fully loaded temperature was less than that of the C3 it is important to note that the safety margin to maximum temperature is much smaller.

Both the P-III at 466MHz and the C3 at 933MHz appeared capable of operating in a low airmovement environment with the Coolermaster passive cooler. However, the mandatory 50% underclock of the P-III and it's small temperature safety margin make its operation somewhat suspect. The C3 operation, although hotter than ideal, worked safely from the start without the need for any modification. It appears that the effort VIA has invested in maximizing power efficiency and reducing heat output reflect well in the C3's real-world performance.

System configurations:

Unique Components	VIA	INTEL
CPU	C3 933 (7x133)	P-III 466 (7x66)
Motherboard	ECS P6VXM2	Abit BE6-II
Northbridge	VIA Apollo Pro 133A	Intel 440BX
Southbridge	VIA 686B	Intel 82371EB
Shared Components
RAM	2x256MB Crucial PC133 cas2 ECC SDRAM
Video Card	ATI Radeon VE
Hard Drive	20GB IBM Travelstar 40GN
Ethernet	3Com 3C905TX

The Radeon VE was chosen because, being based on a mobile core, it produces very little heat and is more than capiable of running with a simple heatsink. Also, it provides multi monitor support as well as a DVI output for digitally driving LCDs.

The hard drive used is a 2.5" laptop hard drive running at 4200rpm. The reason for this is twofold; heat and noise. The laptop drive runs both quietly and cool to the touch. The effects of the slower hard drive could be felt when booting Windows 2000 - it was respectably slower. Once in Windows, the 512MB of RAM keeps Windows from needing to access the hard drive very often, and the speed was not an issue. To use a laptop drive with a desktop drive a adapted kit is needed for cable adaption and mounting.

Benchmarks:

With the exception of the Kernel Compilation Benchmark, all tests were run under Windows 2000, service pack 2.

Sisoft Sandra is a widely accepted synthetic system benchmarking utility. Sandra accurately portrays performance differences between different speeds of similar architectures. However, it is the authors' opinion that when comparing cross-platform results, Sisoft Sandra can show results which simply do not manifest in real-world applications. It is important to keep in mind that these tests are purely synthetic.

For example, the VIA C3 Ezra processor executes x86 load-ALU-store instructions in only one clock as compared to several clocks on other processors.

This instruction is exactly what the Sandra "integer" memory bandwidth score makes use of. VIA's claim of being faster in this area doesn't appear supported even with a 2x advantage in clock speed and memory bandwidth. The Apollo Pro 133A north bridge may be responsible for some or all of this result. Unfortunately the C3 was unable to boot properly when installed in the BX motherboard, so the culprit could not be identified.

Cachemem is a utility used to measure throughput and latency of L1 cache, L2 cache and main memory. Latency measurements test how many cycles it takes to access the first block of memory/cache in a random seek (ie. not burst mode). Cache measurements are purely a CPU attribute, while the memory scores will be directly affected by the chipset.

The C3 holds its own in L1 cache speed, which is good because it has 128kB total of L1 cache. It lags behind in L2 speed, but this is not as big of a problem, since the C3 only has 64kB of L2. This is a strange caching structure, because much (all?) of the data stored in the L2 cache will be duplicated in the faster L1 cache, unless VIA has done something to prevent this (as AMD has, with the Athlon's consolidated 320kB cache).

The memory throughput scores reflect Sandra's results: the P3+440BX at PC66 CL2 beats out the C3+Apollo Pro 133A at PC133 CL2. This is a little surprising, but cachemem simply confirms Sandra's result. This only goes to further credit BX's efficient memory controller.

The Pentium III falls into the standard CPU caching scheme. First, check L1 (3 cycles), then check L2 (7 cycles). This stands to reason because the L2 cache is far larger than the L1, so it takes longer to search. The C3 is running a strange mechanism for caching, because the L2 is smaller than the L1, and it has the same penalty (latency) to check as the L1.

The 440BX flexes its muscle again when it comes to main memory latency. 85 cycles for a main memory access is simply incredible, especially given today's values for DDR systems are over 200 cycles, and RAMBUS PC800 memory tops 350 cycles before the first data is returned!

MadOnion 3DMark 2001 is a multimedia benchmark designed to test DirectX 8.0 and 8.1 3D gaming performance. The ATI Radeon VE is based on the mobile version of the classic Radeon core. As such it does not support most of the DirectX 8.x features being tested and will consequently score poorly. However, comparison between the two platforms, using the same video card and driver, should accurately portray differences in 3D game performance.

Linux Kernel compilation is a test representative of many coding situations, as well as bing a real-world task. A stock 2.4.18 kernel was compiled under RedHat 7.2 using the following setup.
tar -xzvf linux-2.4.18.tar.gz
cd linux
make menuconfig (save and exit, no changes)
make dep
time make bzImage

The returned 'real' time is taken as the result.

Conclusion:

Both CPUs were capable of the assigned task: to run at 100% CPU load with only passive cooling for an arbitrary period of time. The Pentium III running at 466MHz remained faster than the C3 at 933MHz in most tasks, however, speed is not the main concern of this article. There is no way to declare a clear "winner" because both CPUs have their relative strengths and weaknesses.

One thing is certain -- it is very refreshing to use a computer that operates so quietly! Once you have used a system with no fans, and hear long forgotten sounds such as 'birds' or 'your own breathing' the constant 'whirring' of ordinary PCs will start to bother you more and more.

Impressions: