Rambus on the Ropes?

While the smart money won’t be betting against Direct RAMBUS DRAM (DRDRAM), the outlook certainly isn’t as rosy as it appeared a year ago. Despite licensing restrictions that prevents negative comments from manufacturers who have signed on with Rambus, many rumors are being spread that the technology is not all it has been cracked up to be.

The Sure Bet

Since mid-1998, Intel has been extolling the virtues of DRDRAM and vowing to make it the memory standard of the future. As reported by CNET, last October Intel made a $500 million dollar investment into Micron, then followed that up in January with a $100 million dollar investment in Samsung, as reported again by CNET. Sources have indicated that part of the agreement was that these manufacturers could not promote any other form of DRAM over the DRDRAM, in addition to guarantees of supplying Intel with sufficient parts.

Usually, once Intel decides to back something it’s a pretty sure bet that it will become the standard, though not without some grumbling from competitors, and DRDRAM has been assumed to be a shoe-in because of this. As reported in a previous editorial, there were some complaints early on, but there wasn’t much attention being paid to them.

The assumption has been, amongst many industry observers, that DRDRAM would take over the world within a few years, and Intel has been driving that message home with everyone who would listen. Intel and DRDRAM appeared to have won a major market victory when the SLDRAM consortium finally gave up on SLDRAM, as reported in EETimes.

“It doesn’t mean it doesn’t work”

Until earlier this year, Intel maintained that the 820 chipset (code named ‘Camino’) would be shipping by June to support the Pentium III 533MHz chips at a bus speed of 133MHz. Suddenly, in February of this year Intel pushed back the release date of this chipset to September, while denying that there are any problems inherent in the technology.

Though manufacturers are prohibited by their license agreement to make any negative statements publicly, more and more are privately questioning the direction Intel has stubbornly refused to budge from. In fact, Intel even recently backed off plans to provide support for SDRAM on RIMMs as an alternative memory solution, much to the chagrin of smaller systems integrators and OEMs. All of this is pretty scary, and recently some comments have been heard that DRDRAM is very susceptible to errors that ECC won’t fix, and that the interface is far too unreliable to trust with a server’s data.

Even more interesting is the recent story on CMP Net indicating that Intel and Rambus are adding a 700MHz spec to ‘compliment’ the 600MHz and 800MHz specs – supposedly to help DRAM makers improve yields. This seems to give credence to the rumors that until DRAM and chipsets go to .18u technology 800MHz is unachievable. The real question is, what other problems have been encountered that Intel and others have not publicly released because of corporate politics and licensing issues?

SDRAM gains strength

In the midst of all this positioning and maneuvering, VIA and several DRAM manufacturers (including Micron) began working on a PC133 specification (reported in our Industry News on 2/23/99). Even more recently, Enhanced Memory Systems has released their HSDRAM, which is essentially 133MHz SDRAM (though an official PC133 spec is still forthcoming), and has been reviewed on several websites, including Sharky Extreme. Micron will also release PC133 SDRAM as soon as a specification is approved, and several other manufacturers have indicated that they have 133MHz (and even faster) SDRAM parts.

While Intel has been claiming that bandwidth is the main memory issue, the reality is that the majority of PCs (including servers on small networks) are not bandwidth constrained. In fact, only a small percentage of systems are suffering from bandwidth issues. For most systems, latency is the main memory bottleneck, which DRDRAM does nothing to address and may even be worse than current SDRAM solutions.

The main reason that Intel claims that bandwidth is an issue is that upcoming applications, such as voice recognition, full motion video across the internet and other similar ones will tax the ability of servers to handle the data. Intel sees their 600MHz Pentium III processors as the workhorses for these systems, and want to utilize DRDRAM and it’s theoretical 1.6GB/Sec bandwith for this purpose. Unfortunately, at 600MHz (the current maximum speed) DRDRAM will provide 1.2GB/Sec vs. the 1.06GB/Sec that 133MHz SDRAM allows for – hardly a major difference. Estimates have indicated that DRDRAM will be from 15% to 30% (or perhaps more) higher in price than SDRAM solutions

What makes this even more interesting is that once DRAM goes to .18u technology, it is likely not very long before 166MHz SDRAM becomes a possibility, giving a bandwidth of over 1.3GB/Sec – certainly more than sufficient for all but the most demanding of server applications. In addition, with DDR SDRAM solutions the bandwidth effectively doubles, thereby surpassing even the 800MHz DRDRAM solutions by a wide margin. Several manufacturers (including, once again, Micron) are working to have DDR SDRAM available before the end of the year.

So, Which One Will it Be?

Not so long ago, I was attending a conference for mainframers, called SHARE. During one session on Ethernet technology two IBM consultants were commenting about how IBM pushed SNA and Token Ring very hard because of its ‘technical superiority’, though it was also more expensive. The reason that IBM believed in these technologies is that Ethernet is very inefficient (only about 40% utilization is possible), and the (then) 10Mbit bandwidth was insufficient for future needs. These consultants then pointed out that today IBM is fully committed to Ethernet because they finally recognized a simple principle – so many companies had invested in Ethernet because it was cheap that market momentum would encourage engineers to overcome the existing limitations. This was realized when 100Mbit, and then 1Gbit Ethernet was introduced. Now we are talking about 10Gbit Ethernet.

It seems to this writer that perhaps Intel has fallen into the same trap that IBM did. The main difference here, of course, is that Intel currently controls the chipset market and figures they will simply make everyone go to DRDRAM by brute force. Unfortunately for them VIA is making advances in their Slot 1 chipsets, and will be supporting 133MHz SDRAM long before Intel releases their 820 chipset. Additionally, VIA already has DDR support in all of their recent chipsets (as well as ESDRAM).

With all the delays, discontent amongst manufacturers and market fear of an Intel controlled world, we may see Intel forced to concede that DRDRAM will only fill a niche market. Once a standard is established, it becomes very difficult for any company (except one that has a monopoly) to change the momentum – especially when the ‘replacement’ standard is more expensive. Intel has usually been the one to establish the standards first for this very reason, but this time it seems possible that they backed the wrong horse…Of course, I could be wrong, too.

Be the first to discuss this article!