Page 1:Bringing Pentium 4 To The Masses
Page 2:Pentium 4 Chipsets
Page 3:Pentium 4 Chipsets, Continued
Page 4:VIA's P4X266 Chipset
Page 5:Testing Procedure
Page 6:Quake 3 Arena 640x480
Page 7:Quake 3 Arena 1024x768
Page 8:AquaNox AquaMark
Page 9:Unreal Tournament
Page 10:MPEG4 Encoding With FlasK MPEG And DivX
Page 11:Conclusion - Pentium 4 Without Rambus Is Very Attractive
Pentium 4 Chipsets, Continued
Intel's upcoming 'low-cost' solution for Pentium 4, the i845-chipset, will initially only support PC133 SDRAM, which cuts its available memory bandwidth down to a mere 1033 MB/s. We have shown in a previous article that this low memory performance slows down Pentium 4 quite badly. By the end of this year or at the beginning of 2002 Intel will also allow i845 with DDR-SDRAM. This will double its memory bandwidth and should finally enable a reasonable performance of i845 platforms. The official reason for this delay are validation issues, as Intel seems rather displeased with the current chaos in the DDR-memory market, which is missing tight specifications that would ensure reliable operation of each DDR-SDRAM module (Intel: "Neither AMD nor VIA tried to do anything about it"). There's also still the old Rambus conspiracy theory, which claims that Intel is unable to supply i845 with DDR-support because of license agreements with our friend Rambus Inc.
VIA's P4X266 supports any kind of SDRAM or DDR-SDRAM. Once teamed up with PC2100 DDR-SDRAM, the available memory bandwidth is with 2133 MB/s still lower than what Intel's 850/dual-channel RDRAM solution is able to offer. It is rather questionable though, if the mere memory bandwidth number is indeed able to tell the whole story. Most applications require low latency as well, and that is where RDRAM hasn't been able to outdo SDRAM or DDR-SDRAM solutions so far.
Here's a table that shows the memory bandwidth numbers of the different memory types.
|Speed Grade||Memory Type||Clock||Clock Edges Used||Channels||Bus Width||Peak Bandwidth (clock x edges x bus width x channels)|
|PC100||SDRAM||100 MHz||1||Single||64 bit||800 MB/s|
|PC600||RDRAM||266 MHz||2||Single||16 bit||1066 MB/s|
|PC133||SDRAM||133 MHz||1||Single||64 bit||1066 MB/s|
|PC150||SDRAM||150 MHz||1||Single||64 bit||1200 MB/s|
|PC166||SDRAM||166 MHz||1||Single||64 bit||1333 MB/s|
|PC800||RDRAM||400 MHz||2||Single||16 bit||1600 MB/s|
|PC1600||DDR SDRAM||100 MHz||2||Single||64 bit||1600 MB/s|
|PC600||RDRAM||266 MHz||2||Dual||2 x 16 bit||2133 MB/s|
|PC2100||DDR SDRAM||133 MHz||2||Single||64 bit||2133 MB/s|
|PC800||RDRAM||400 MHz||2||Dual||2 x 16 bit||3200 MB/s|
|PC1600||DDR-SDRAM||100 MHz||2||Dual||2 x 64 bit||3200 MB/s|
|PC2100||DDR-SDRAM||133 MHz||2||Dual||2 x 64 bit||4266 MB/s|
The table shows quite nicely, that of all single-channel memories, PC2100 DDR-SDRAM is the one with the highest bandwidth. However, we should not forget that one SDRAM or DDR-SDRAM channel requires 64 data lines, while RDRAM only requires 16. That is why a dual-channel Rambus solution requires 32 lines only, which is easier to implement, as long as there aren't any problems to deal with the very high memory clock of 400 MHz. The dual-channel DDR-SDRAM solutions listed at the bottom of the chart are currently only realistic with NVIDIA's nForce chipset. The nForce chipset as well the nForce reference boards show that dual-channel DDR-SDRAM is indeed doable. A dual-channel PC2100 DDR-SDRAM solution beats the bandwidth numbers of all other memory configurations.