Our last benchmarking of the Intel Pentium 4 based platforms in the SPEC CPU2000 took place in February 2002. That time we used several Intel's chipset to find out how the test depended on different parameters - FSB clock speed, memory frequency and type.

Since that time there were a lot of new chipsets, but the real breakthrough happened only in the end of 2002 with the advent of the Intel E7205 with the dual-channel DDR SDRAM memory controller. SiS 655 chipset has recently joined it.

So, today we will draw one more intermediate conclusion before the Pentium 4 with the FSB 800 MHz and Intel's new i875 and i865 chipsets start spreading out.

Testbed configuration

The testbed is based on the Intel Pentium 4 3.06 GHz. All of the tests are carried out under the Microsoft Windows XP Pro SP1. 

Platforms:

• Intel i850E (ASUS P4T533) 
• Intel i845PE (Gigabyte GA-8PE667 Ultra) 
• Intel E7205 (MSI GNB Max) 
• SiS 655 (Gigabyte GA-SINXP1394) 
• SiS R658 (ABIT SI7-G) 

Each platform was coupled with a respective memory type - RDRAM for i850 and 658, DDR SDRAM for the rest. A bit later we will give you some more details on the memory configuration.

SPEC CPU2000 configuration: Intel C/C++/Fortran 7.0 compilers, Microsoft Visual.NET (160103) libraries. SSE2 optimization.

First of all, let me focus on the peculiarities of the systems tested and specify their possible memory configuration. 

The i850E is Intel's classical chipset for the Pentium 4. Originally it was the only chipset (or, rather, its predecessor - i850) that could ensure the memory throughput required for the processor. The throughput of the dual-channel PC800 RDRAM corresponded to 3.2 GB/s of the 400MHz QP bus of the first Pentium 4. The i850E currently supports both CPUs with the FSB 533MHz and PC4200 memory. So, the data rates are balanced again (it refers only to the CPU/memory tandem; there are other consumers of the bus but they don't contribute much to the SPEC CPU2000). For the tests we installed Samsung's 32bit 512MB PC1066 module.

The i845PE chipset is a mainstream solution for Socket478 processors. It supports a maximum of DDR333 the speed of which is noticeably lower than 4.2 GB/s required by modern processors. The i845PE chipset was coupled with two TwinMOS DDR-400 256MB modules working as PC2700 with the timings 2-5-2-2.

Intel's E7205 was meant to bridge the processor/chipset data rate and chipset/memory one. Two channels of the DDR266 ensured the overall throughput of 4.2 GB/s. For the tests we installed one and two TwinMOS PC3200 modules. At 266 MHz they had timings 2-5-2-2.

Dual-channel SiS 655 chipset for DDR memory was tested with the same TwinMOS modules. It provides the maximum throughput of 6.4 GB/s. It's 1.5 times more than the Pentium 4 with 533MHz bus needs. Since the SPEC CPU2000 is a narrow-range benchmark, we might not notice those 6.4 GB/s. But our last year's tests of the AMD Athlon XP platforms showed that the reserve of the memory throughput can be useful in such tasks as well. The tests were carried out at the memory clock speed of 266, 333 and 400 MHz. In the first two cases the timings were 2-5-2-2, in the third one - 2-6-3-3. Note that this chipset has three modes for the memory controller:

• one channel, 64bit bus (standard); 
• one channel, 128bit bus; 
• two 64bit channels. 

In the first case you can use any number of the modules, and in the other cases two (or four) identical modules must be installed.

SiS R658 is the first product of this company for RDRAM memory. In spite of the official compatibility with PC1200 (PC4800), the speed in this mode doesn't differ much from the PC1066. Taking into account that the mode worked as an overclocking function rather than a standard one, I decided to use the same Samsung PC1066 module in the tests. It's quite possible that next BIOS versions of this board will be improved.

All the mainboards tested proved to be reliable in the tasks used  - none of them hung during the tests.

All the performance settings in the BIOSes were set to maximum. In our next reviews you will find more details about influence of the DDR memory's timings.

Tests

Let's start with the integral diagrams SPECint_base2000 and SPECfp_base2000. Further we will choose most interesting groups and take a closer look at them.

The i850E remains a leader though the others are much closer now. We can see again that the combination of the Pentium 4's QP bus and RDRAM memory is very beneficial. Besides, the developer does matter - SiS's solution loses both to the i850E and to the i845PE.

Secondly, we can see that the Intel's chipsets outscore their competitors in spite of the same memory configuration. This proves again that Intel's products are more advanced.

Besides, the performance of the SiS655 in the "2x64" mode is better than in the "128" one. And the system with two DDR400 memory modules works slower than those equipped with two DDR333. At 200 MHz the modules have greater latency than in the DDR333 mode, and the dual-channel DDR333/400 is wide enough for the 533 MHz QP bus of the Pentium 4.

The iE7205 with the dual-channel DDR266 performs pretty well. The SiS655 with its dual-channel DDR333/400 does not fall much behind them.

Finally, the dual-channel memory controller allows for a 6-10% growth, which is not that bad.

There is an upper limit around 1100. It can be reached by the most efficient systems.

Now take a look at the scores in SPECfp_base2000.

In general, the scores are close to the first diagram, only the data scatter to a greater degree, and the SiS655 performs better. 

• The i850E based board makes every effort to remain the leader; 
• SiS655+DDR333 goes on a par with the i845PE+DDR333; 
• Dual-channel DDR400 on the SiS655 outedges the dual-channel DDR333; 
• In the dual-channel mode the iE7205 markedly loses to the SiS655 with the DDR333/400 memory; 
• The dual-channel configurations outdo the classical ones by 15..24%. 

There is an upper limit as well at 1100.

Now we are going to closely examine some groups.

The first pair is i845PE+DDR333 and SiS655+DDR333. Let's see which subtests induce the difference in the integral scores. Note that the DDR333 is not enough for the Pentium 4 3.06 GHz, and these tests are focused mostly on the memory controller in the chipsets studied.

In the CINT2000 the breakaway of the i845PE reaches 10%; it takes place primarily in the memory-dependent subtests. In the CFP2000 the chipsets win up to 6% from each other in different subtests. The integral scores are equal at that. The SiS655 wins in the memory-hungry 171.swim and 179.art. This test shows that Intel is not the only company able to make fast memory controllers.

By the way, the DDR400 doesn't let the SiS655 outrun the tandem of i845PE+DDR333 in the CINT benchmark (the lag is 2.19%). But thanks to the 171.swim, 172.mgrid and 179.art the SiS chipset outpaces it by 5.14% in the CFP2000.

The second pair is the SiS655 chipset with different memory configurations. The precise data do not fit in the diagrams. The brighter colors refer to the dual-channel configurations.

These diagrams do not reveal anything new. There are only several aspects I'd like to focus on:

• Replacement of the single-channel DDR266 memory with the dual-channel DDR400 (or DDR333) allows for a considerable growth - up to 58%. The integral scores grow up by 18 and 40% respectively for SPECint_base2000 and SPECfp_base2000. 
• The dual-channel configuration with the DDR266 scores pretty low results. Although its throughput corresponds to that of the 533MHz QP bus of the Pentium 4, the actual speed is approximately equal to the single-channel DDR333/400. On the other hand, two channel are the most beneficial on the DDR266. For the SPECint_base2000 it makes 10.40/8.13/5.95% (DDR266/333/400), and for SPECfp_base2000 it's 23.82/20.00/14.45%. 
• Speeds of the dual-channel DDR333 and DDR400 are approximately equal, and they are noticeably higher than that of the rest (in the memory-hungry tests, which include are except 177.mesa and 200.sixtrack for the CFP2000). 

The last couple is i850E+PC4200 and SiS655+DDR400(2x64).

The two platforms have very similar performance though the manufacturer, chipset and memory are different. The SPEC CPU2000 benchmark finally justifies its name of J. But if you look at the precise data, the picture won't seem so smooth: 

Have a look at some results:

• 164.gzip & 256.bzip2 - despite the same class of tasks, 164.gzip works faster on the SiS, and 256.bzip2 on the Intel. 
• 176.gcc and (175.vpr, 181.mcf, 300.twolf) earlier demonstrated a strong dependence on the memory clock speed but the SiS wins in the 176.gcc and loses about 4% in the other tests mentioned. 
• The picture is similar for the 171.swim and 179.art. 

Conclusion

For the spring 2003 the i850E chipset coupled with the RDRAM memory is the most efficient system for the Intel Pentium 4 with FSB 533 MHz. But its advantage is so intangible that the modern chipsets with the dual-channel memory system for DDR SDRAM almost catch up with it. Most probably, the situation will change with the advent of Intel's chipsets of the i865/i875 series supporting dual-channel PC2700/PC3200 memory, and the tandem of i850+RDRAM will lose the palm once and for all.

Intel's first dual-channel chipset E7205 is a good way to lift up the RAM bus speed by using two channels for the DDR SDRAM memory. But taking into account that the i875 is already launched, the iE7205 can be considered an obsolete model, especially because it supports only DDR266 and processors with FSB 400/533 MHz.

The classical i845PE is Intel's basic version for the DDR266/333 memory. It shows a good speed when coupled with the DDR333 and maintains its leading position in this class. Some manufacturers consider the i875 too expensive and, therefore, overclock the i845PE for support of the new processors with the 800 MHz FSB and DDR400 memory. That is why this chipset is really competitive, and we will soon estimate its operation in non-standard modes.

SiS 655 chipset performs well, and at the moment it's the most efficient solution for the dual-channel DDR333/400 memory. It almost catches up with the i850E when coupled with fast high-quality memory.

The results of the other SiS chipset - R658 - have disappointed us, because it works like the dual-channel DDR266. However, the situation can change with the new BIOS versions, especially because the SiS655 got scores comparable to the Intel's products.

In general, the manufacturers were right when they moved to the dual-channel chipsets for DDR SDRAM. Especially on the threshold of the processors with a 800MHz bus. The DDR333/400 memory support is useful as well especially because the memory modules which ensure operation with timings like good products with DDR266 are already available.

The practical conclusion is that if you need to compare processors, their "framing" (surrounding components) should not affect the outcome and should be as fast as possible.

Write a comment below. No registration needed!