For quite a long while already, since the times of 486 processors, the PC system bus has been falling more and more behindthe constantly accelerating CPUs. It was exactly at that particular moment that Intel for the first time refused to make itsprocessors work at the frequency synchronized with that of the system bus, and applied the method of FSB multiplying. Thisfact got reflected even in the name: 486DX2. Although the system bus frequency remained the same, the CPU performance nearlydoubled.

Further on the difference in clock frequencies of various system components kept increasing. While the system bus frequencyrose at first up to 66MHz and then to 100MHz, the PCI bus retained its 33MHz, AGP bus stayed at the standard 66MHz, etc. As forthe memory bus, it remained synchronized with the system bus (the name required it: Sybnchronous DRAM, SDRAM). This is how PC66specification appeared and then also PC100, and PC133, which required a bit more effort this time.

However, while the memory frequency became only 1/3 higher and hence its bandwidth also rose by 1/3 (from 800MB/sec to1,064MB/sec), the processor frequency managed to jump up by two and a half times: from 400MHz to 1GHz. Doesn't it seem a bitmisbalancing to you? The bandwidth of PC133 SDRAM makes only 1,064MB/sec, while the today's PCs require at least 1GB/sec fora CPU supporting 133MHz FSB, the same 1GB/sec for AGP 4x bus, 132MB/sec for 33MHz PCI bus. Altogether this makes 2.1GB/sec.See the misbalance? Over 2 times!

But the today's equipment of most memory manufacturers, does not allow them to further increase PC133 SDRAM frequency:166MHz SDRAM turns out too expensive, especially taking into account the today's requirements for system memory. As youremember, this particular thing if disregarded, resulted into a failure for Direct Rambus DRAM. At the same time, thereare very cogent arguments in favor of memory bus and system bus synchronization.

The technologies called to improve SDRAM by means of the additional SRAM cache, like ESDRAM, or by means of the workoptimization, like VCM SDRAM, didn't help at all. In fact, a very popular technology came to rescue: simultaneous datatransfer along both signal edges, when two data packs are transferred within one single clock. In case of the today's64bit bus it makes two 8-bit packs (16Bytes per clock). Or in case of our 133MHz bus - 2,128MB/sec instead of 1,064MB/sec,which is exactly what all modern PCs need.

As for the price, it hardly differs a lot from that of the ordinary 133MHz memory: the technology is absolutely the same(including TSOP chips packing technology and not microBGA as by RDRAM), the equipment is also the same (power consumptiondoesn't differ practically from SDRAM manufacturing, and the chip surface is only several percents bigger). This uniquecombination of affordable cost and appropriate performance turned pretty interesting for pragmatic DRAM industry in thefirst place. They used to decide on PC66, PC100, PC133, etc. in the same way…

However, these specifications, which had the memory bus clock frequency even in their names, as well as Direct RambusDRAM, which is based on the frequency equal to clock frequency multiplied by the same two packs per clock as we saw by DDRSDRAM - PC600, PC700, PC800, didn't seem attractive enough to the companies dealing with DDR SDRAM development. The latter,and to be more precise, the marketing departments of the latter, chose the method, which could provide them with the highestnumber in the names. Namely, they took the peak bandwidth and as a result got PC1600 for 100MHz and PC2100 for 133MHz DDRSDRAM chips.

In fact, this naming system has been invented quite recently, although DDR SDRAM chips have been manufactured for a longtime already: 64Mbit chips appeared for the first time about 2 years ago in mid 1998. By that particular time, in December1998, when Intel had been supporting RDRAM for quite a while already, a new open DDR SDRAM specification, which didn'trequire any license payments from the manufacturers using it, got finally approved. The main adherents of this newspecification, like in case of PC133 SDRAM, appeared VIA and IBM. By that time they had already clearly defined theirstrategy directed towards architectures alternative to RDRAM. A few months later, in May, the specification of 184-pinDIMM modules also got approved and the development of DDR SGRAM came to an end.

About a year and a half later DDR SDRAM grew ready for commercial manufacturing: 133MHz 64Mbit DDR SDRAM chips samplesmeeting PC2100 spec were well prepared for large-scale release.

However, memory module manufacturers were not the first to use DDR chips. At that time graphics cards manufacturers turnedout in a more favorable position: they could use whatever they wanted on their cards to get a standard output signal. Besides,memory bus bandwidth has always been a bottleneck for graphics cards rather than for CPUs. That's why graphics cardsmanufacturers made effective use of DDR SDRAM/SGRAM support much earlier than their CPU colleagues. Hardly had a fewmonths passed after the appearance of the first chip like that, GeForce 256, when the first cards with DDR SDRAM andSGRAM chips on boards cropped up.

The chips of the first group of DDR cards supported mostly 150 and 166MHz (the end-frequency - 300 and 333MHzcorrespondingly, the bandwidth of 128bit bus - 4.8 and 5.2GB/sec). We have every reason to suppose that autumngeneration of the graphics chips will base on 183MHz chips (366MHz, 6GB/sec), and in 2001 we will witness masslaunching of the graphics cards with 200MHz (400MHz, 6.4GB/sec).

The fact that SDRAM/SGRAM got replaced by a twice as fast memory didn't take long to tell. The performance ofthose graphics cards, which appeared in systems with a powerful CPU increased by almost 1.5 times in applicationsloading exactly the memory bus (such as 32bit color, for instance).

Taking into consideration the already available information on the developers' plans for the coming year, wehave every right to announce DDR's indisputable victory over the notorious RDRAM. Having successfully pushed forwardAGP with its i740, Intel luckily gave up all further attempts to exert influence in this field, and now it's the marketthat controls it. Expensive RDRAM turned out absolutely undemanded, especially since 128bit memory bus makes DDR SDRAMeven cooler than a dual-pipeline RDRAM.

As for DIMM DDR SDRAM memory modules, the situation is a bit different: nobody seems to really need them now, becausethey require special mainboards built on the chipsets supporting this memory type. We expected the first user chipsetsupporting this memory type from VIA in autumn 1999, and then in winter 2000, and then in spring… But now it looks asif there is not so much waiting left. In Q2 we should be able to see the first VIA chipset supporting DDR SDRAM - ApolloPro266.

Besides the already familiar 133MHz FSB and AGP 4x mode, there will also appear DDR SDRAM support, as well as some other enhancements.Moreover, VIA Apollo Pro266 will also support dual CPU configuration,which has already become an official feature of Apollo Pro133A.

A bit later in Q3 we can witness the launching of Apollo Pro266 version with the integrated graphics core - PM266. Unlike PM133,equipped with a relatively weak for the third quarter Savage4, this chipset will boast an integrated Savage2000 (GX4C).Its performance should be more than enough for low-cost systems, which are currently the main market niche for integratedchipsets.

Besides, we can't help mentioning the server solutions with DDR SDRAM supprot, which are most likely to appear by the endof the year, because some manufacturers are known to be currently developing solutions like that. Take for instance ServerWorkswith their new chipset.

The second half of the year is also already planned for the launching of Athlon DDR chipset - KX266, which should be verysimilar to its Pentium III fellow - Apollo Pro266. But just in case AMD decided to make sure that everything would go tiptopby launching its own chipset with DDR support in Q3 - AMD 760. It is expected to support a new EV6 system bus frequency -133MHz (266MHz), and of course, 133MHz PC2100 DDR SDRAM and ATA100. Soon after the launching of AMD 760, they are going tointroduce their multiprocessor AMD 770 with similar features.

Since we came to speak about multiprocessor chipsets intended for server platforms, we can't help mentioning two moreparticipants of this competition: Samsung with their Caspian developed in cooperation with AMD and ServerWorks with theirServerSet family, which should acquire a DDR SDRAM chipset for Intel processors in the first half of this year already.

Taking into account such factors as the cost of RDRAM, the performance difference between RDRAM and DDR SDRAM and theperformance drop of RDRAM memory subsystem with the growth of memory size, the overwhelming majority of server manufacturersare going to prefer DDR SDRAM to RDRAM. Even Intel has to accept this state of things that is why they are planning to providetheir next server chipset for x86 (i870) with DDR SDRAM support. Moreover, Intel won't be the only one in the market willing tosupport DDR. Besides the independent developers, DDR support will attract most server makers, such as IBM, NEC, etc., who alsodevelop chipsets for their systems.

In a quarter after the launching of corresponding chipsets, we can expect the corresponding mainboards built on them to comeout. So, the first mainboards supporting DDR SDRAM memory modules should be available in Q3 2000 already. This is exactly thedeadline set by most mainboard manufacturers. The first and the only inconvenience for those who will decide on a mainboard likethat will be a new DIMM modules form-factor.

Unfortunately, nothing is free in this world that is why modules form-factor had to be changed in order to double the memorybandwidth. Although the modules remained of the same size, the number of pins grew from 168 to 184. And since the notch key is movedto a new location, you won't be able to install these DIMM DDR SDRAM modules into the today's slots.

Well, let's pass over to a more interesting thing: real performance values. Unfortunately, we can't say anything about theprobable results of VIA chipsets. What we know is the performance of Samurai from Micron. By the way, VIA licensed some Micron'sachievements on DDR, and as for Micron, they don't even plan to enter the chipset market considering Samurai a spare versionstimulating the sales of DDR SDRAM chips together with the registered 133MHz DDR SDRAM modules (server version - more reliablebut slower DIMM modules, user version - faster, non-buffered DIMM modules).

Tests made by InQuest Market Research

StremD is a test recognized by the memory industry, which evaluates the efficient bandwidth of the memory bus. Take alook at the results: hardly any commentary is needed. The results give us a clear idea of all the further results obtainedin real life applications. Of course, these applications do not depend so greatly on the memory bus bandwidth. That is whythe differences between RDRAM and DDR platforms are so smooth, but anyway, in real life applications DDR proves a bit coolerthan RDRAM.

And now a few words about the near future. DDR SDRAM DIMM modules standard implies using up to 200MHz chips providing 400MHzend frequency and the bandwidth of 3.2GB/sec as we saw by dual-pipeline Direct Rambus DRAM. By the time DDR SDRAM totally exhaustsitself, namely about 2003, DDR-II should be already getting started.

DDR-II chips are expected to work at 100MHz and up, but since they also transfer 4 data packs per clock their bandwidth will make3.2GB/sec, too. With this data transfer method (32Bytes per clock) the performance growth of DDR-II will reach its maximumwith the clock frequency increase. We mean fourfold growth here: 150MHz will result into 4.8GB/sec, 200MHz - 6.4GB/sec.

The modules on these new chips, as well as the modules on DDR chips, will also feature their own form-factor (230 pins)and require new chipsets. The same thing is true for Advanced DRAM Technology chips due about the same time.

Well, there are still three years left until then that is why for three years we will have to decide between DDR SDRAM andDirect Rambus DRAM. If Intel doesn't try to force the market in favor of the latter (and it undoubtedly does!), the outcome ofthis competition is more than evident: DDR SDRAM takes the lead. Otherwise, the situation may take a really unexpected turn,because it's pretty hard to say what will tip the balance: Intel's financial power or industry's common sense, and in whatproportion to each other they will appear.

Anyway, if we try to veer away from all extremes, we will have the right to state that DDR SDRAM seems to have a morepromising future than a year ago, no matter what happens. This year was very rich in different events: AMD launched itsAthlon, strengthened its positions in the market and VIA staked DDR SDRAM. So, no matter what happens to Intel solutions,those users who will decide on AMD processors till the end of this year are already destined for using DDR SDRAM. And infact, this is not bad at all, especially if the price on RDRAM doesn't change drastically by the end of the year. In otherwords, this will be a good argument in favor of AMD/VIA solution if you are usually guided by common sense and not byadvertisements when purchasing hardware.

ServerWorks plays a guarding role for Intel today. They are intended to help Intel in all those parts of x86 server market,where Intel is unable to perform. The platforms from ServerWorks will be a solid response to i840 with dual Rambus pipelines inthe server and workstation market (like VIA chipsets are to the ordinary users PCs).

According to the preliminary benchmarks for Samurai, the system built on it performs as that on i840 and sometimes proves evenfaster. Bearing in mind the today's RIMM cost, which is very unlikely to get a few times lower this year, as well as the memorysize of servers and workstations, we get a really impressive price difference between the solutions with DDR SDRAM and RDRAM.Though the performance of both systems is practically the same, the price differs by thousands of bucks!

The conclusion will sound as follows: DRAM manufacturers cannot give up DDR SDRAM. The market for this memory type is quitesignificant and besides, switching to this memory type, won't require any huge extra expenses. The chips manufacturing costsalmost the same as that of SDRAM chips of the same clock frequency. RDRAM in its turn costs so much that the users cannotafford to acquire more memory for their PCs even if the demand is quite high. So, we come to a paradox: once we disregardPC133 SDRAM as an older generation technology, the market will turn out unable to supply any system memory. Do youreally think we have the right to consider a crazily expensive RDRAM an acceptable solution? In fact, it could have beenan awful mistake not to use this opportunity: DDR SDRAM was simply inevitable.

The other part of the picture concerning chipsets and mainboards is also getting clear. In the second half of the year 2000there will be quite enough solutions in the market with DDR SDRAM support to cover the entire area: VIA and AMD chipsets - High-EndPC based on Pentium III and Athlon, AMD and Samsung chipsets - servers and workstations based on Athlon, ServerWorks chipsets -servers and workstations based on Pentium III.

Available, cheap, providing RDRAM performance for SDRAM price... It will survive. And not just survive; it will have a coollifetime...