Battle of the Behemoths
Over the past year, we have watched CPU speed announcements coming at a regular basis from chip-makers AMD and Intel. AMD, who is having a good year in both sales and production, has announced 1.0 GHz CPUs, while Intel has announced a 1.13 GHz Pentium III in “limited quantities,” which roughly translates into “you can’t have it.” This may seem disappointing to those who want the top of the line. However, not too far below, you can find the Pentium III 933MHz, the current “speed king” as far as Intel’s mass-marketed CPUs go.

The Chip
The Pentium III 933MHz is identical to both the PIII-1.13GHz and its lower clocked 500-866MHz siblings. They share the same 28-million 0.18 micron Coppermine core, and are physically identical in every way.

Coppermine-core based Pentium IIIs have 32KB of on-die L1 cache running at clock speed, the same amount as has been in all Pentium III CPUs. Where Coppermine’s differ however, is that rather than 512KB of off-chip four-way associative L2 cache that ran at 1/2 clock speed (Thus making the L2 cache of a PIII-450 run at 225MHz) across a 64-bit bus, the Coppermine core has 256KB of on-die 8-way associative L2 cache that runs at the full speed of the processor, across an increased 256-bit bus. These changes are the feature Intel calls “advanced transfer cache.”

By moving the L2 cache on the chip, as it is in the Celeron line of CPUs, Intel allows for increased L2 Cache performance that well makes up for the halving of the amount of L2 cache. L1 and L2 cache store information for quick retrieval in order to stop the CPU from hitting the slower SDRAM when looking for data.

Thanks to the 0.18 micron Coppermine core, the PIII-933’s core voltage is just 1.65v, down from 2.00 in the Katmai core Pentium IIIs. By using the smaller 0.18 micron manufacturing process, the die-size of the chip shrinks 17% from 128mm^2 to 106mm^2 even with the addition of on-die L2 cache. This smaller manufacturing process allows the core to run cooler, and to run at higher clock speeds than previously possible. The size of the core has gone up from 9.5 million transistors in the Katmai core PIIIs to 28 million, due to the addition of L2 cache to the core.

CPU Specs:

The Tests
In order to benchmark this chip, I tested it on two motherboards, the ABIT BX133 and the Soyo 7-ISM. The ABIT, while using an older BX chipset, has the ability to adjust the CPU voltage, allowing for easier/more successful overclocking. The SY-7ISM has less overclocking capability, because you may only adjust the bus speed and the SDRAM speed (100MHz or 133MHz). While allowing you to use PC100 SDRAM, this does not help your ability to overclock. Thus, the SY-7ISM was only successful at 933 and 980MHz, while the BX133 was successful at 933, 980 and 1015MHz.

System Specifications
Pentium III 933EB MHz CPU provided by Intel.
128 Megs PC133 Memory
Western Digital 7200RPM 10.2 gig IDE Hard Drive
Sound Blaster Live! Value Sound
3Com 3C905TX-C PCI 10/100 NIC
300 Watt AMD-Approved ATX Power Supply
The above components were purchased from Specialty Tech.
3DFX Voodoo3 3000 16Meg AGP
Operating System: Mandrake 7.1 with 2.2.15 Kernel.

Kernel Compiles
To test the CPU, I ran a variety of benchmarks on it. The first benchmark is a set of timed Linux kernel compiles. Compiling a kernel is a common action for a Linux user, making it a very valid benchmark for a Linux system. To do the testing, I used a Mandrake 7.1 installation. I configured the kernels by typing “make config” and selecting the defaults (holding down enter works nicely). I type “make clean; make dep; time make bzImage” in order to time the kernel compiles.

As you can see, the CPU performs better at kernel compiles under the i815 chipset-based SY-7ISM than on the BX-based BX133. Since the i815 is optimized for use with 133MHz front side bus Pentium III CPUs, and thus provided better results.

Distributed.Net
Next, I used the Distributed.Net client to perform benchmarks using “dnetc -benchmark.” This benchmark stresses the CPU, using CPU-specific engines. So, while not fair for cross-platform comparison, it is valid when comparing similar CPUs, or in this case, the same CPU at different speeds, on different motherboards.

Here, we find that the PIII runs slightly better on the BX-based BX133 board, giving a slight increase in speed over the i815 based 7-ISM.

HDParm
Using HDParm measures the performance (cached and un-cached) of hard drives under Linux. While more an issue with motherboard and drive speed, it does scale with CPU speed, and so is included.

As you may notice, the SY-7ISM outperforms the ABIT BX133 greatly in cached hard drive performance. It should be noted that this is because the primary IDE controller on the ABIT BX133 is UltraDMA33, while on the SY-7ISM the primary IDE controller uses UltraDMA66, doubling the amount of bandwidth available to the drives, thus allowing the cache to transfer at a higher rate. This only affects cached speed and in normal desktop usage, does not reflect greatly increased hard drive speed under normal desktop conditions, but would help with file servers and other uses like that.

Overclocking
The PIII-933 is an OK overclocker, because it is limited by a 7.0 multiplier. In order to get it to work at higher speeds (such as 1015MHz), you may need a board that allows you to increase the core voltage of the CPU, such as the BX133, as simply increasing the bus speed doesn’t always produce reliable results. This isn’t a CPU you’d buy to overclock, however. Most overclockers are looking to save some money, and you won’t do that by trying to overclock one of Intel’s highest-clocked CPUs.

Conclusions
The Pentium III 933 is a good CPU, and would be quite nice in a single or dual-processor workstation. However, you can save nearly $200 by going with an AMD 950MHz CPU if you are interested in a single-processor system. If you lean towards Intel CPUs, or are looking to build a multi-processor workstation, the Intel Pentium III may very well be the way to go. Last, in the interest of full disclosure, I feel I should reiterate the CPU used in these tests was provided by Intel, but should perform identically to a retail Intel CPU. The BX133 and SY-7ISM were provided by ABIT and Soyo, respectively.