Hi all; An interesting link. How Dataquest saw Rambus a year ago:
Dataquest, Tactical Memories Newsletter, Vol. V, No. 17
Google cache, November 27, 2000
1. Shedding Some Light on the Rambus Question
You have to feel sorry for Rambus. Rather like a flawed genius, the technology's ability to produce a high level of performance is unquestioned, but the company appears to have an attitude problem.
After years of promoting the Rambus standard as Intel's memory technology for the future, Intel CEO Craig Barrett publicly renounced the company in the "Financial Times."
Unsurprisingly, given the commercial importance of future PC memory technology to several interested parties, namely those in Figure 1 -- Intel, Rambus, DRAM vendors, PC OEMs and Microsoft -- the Rambus debate, more than current DRAM pricing levels, is now the key issue affecting the outlook for the DRAM industry.
The nub of the issue is this: Could RDRAM really be rejected as a mainstream PC memory technology and, if so, what then?
Dataquest Analysis
The sad thing about this whole Rambus saga is that it was entirely avoidable, and all the interested parties have to bear some responsibility for that.
Although RDRAM has made very little penetration into the mainstream PC marketplace thus far, the introduction and ramp of Pentium 4 (and its associated RDRAM-only chipset) over the next year ought to have been enough of a catalyst to kick-start RDRAM volume production. Most of the technical and logistical obstacles have been overcome with the last remaining barriers to mass adoption, namely unit cost and price likely falling as volume RDRAM production increases.
However, Pentium 4 is not expected to hit the mainstream commercial PC space until the end of 2001 or early 2002 at which time "Brookdale" (the Pentium 4 chipset with SDRAM support) is expected to be available. In effect, this could give enterprise users the choice between RDRAM and SDRAM systems. So which would they likely choose? Obviously, whichever is cheaper, and this is the fundamental problem for RDRAM. Without a commitment to RDRAM technology, volume production is unlikely, and without volume production unit costs, prices will be slow to decline.
From a user perspective, there are two main arguments against RDRAM.
The "We Don't Need the Increased Performance" Argument
Mainstream enterprise users argue they do not need the performance of RDRAM and would prefer to go with a cheaper and less risky alternative, such as SDRAM or DDR. If Intel sticks to its decision to go with its Pentium 4 insurance policy and develop an SDRAM chipset, it could be seen to be endorsing this route for mainstream PC users despite assurances to the contrary. The problem with this solution is that it is inherently short term.
The performance argument may hold for now but, historically, satisfaction with any level of performance does not last forever. DDRDRAM technology has the potential to run out of steam compared with RDRAM, which offers more scalable performance. This is important because big, lethargic software is most likely to become bigger and slower as time goes on -- the buying public upgrades for added features and not for performance. As long as the hardware's speed continues to improve, Microsoft will be able to devote its programming resources to adding features rather than to streamlining performance. It is a win-win for both Intel and Microsoft if hardware performance continues to improve, and consumers support that behavior. Additionally, Intel staunchly affirms that RDRAM bus bandwidth is an absolute requirement to support anything faster than a 1.2GHz processor.
There is a case for choosing DDR to improve system performance in the short term but it can add to system cost and for that reason should not be considered a major candidate for PC main memory in 2001. There are too many technical problems with DDR that have not yet been sorted out to allow it to find its way into main memory in 2001. In a nutshell, no two vendors' DDRDRAMs are completely compatible with each other's, and the details of making an economical motherboard that supports DDR have yet to be worked out. Having said that, PC graphics is a completely different issue and DDR will take over that space very shortly because the technical issues are much less difficult on a graphics card. Graphics cards are tightly controlled environments with point-to-point connections between the DRAM and the graphics controller chip. Likewise, servers can afford eight-layer motherboards, which can kill the noise that would plague a designer of a four-layer board for a PC. [Bilow: Obviously Datquest was very wrong about the number of layers that DDR would require.]
The "It Is Too Expensive" Argument
Now to DRAM cost. Everything is relative. The cost of a 64Mb SDRAM two months ago was $8.00. The cost for it now is $4.00. How much will it be in 2001? Down to $2.00 or back up to $8.00? Whichever way you look at it, when you compare prices for the same device over time you can come up with 50 percent reductions or 100 percent premiums. It is all to do with supply and demand. Of course RDRAM is more expensive to make now but that is because volumes have not ramped. RDRAM will always have a higher silicon area than a comparable density non-RDRAM device but that would be an irrelevance if RDRAM was produced in much higher quantities.
Those of us with long memories will recall that the die area penalty of a 4Mb SDRAM over the 4Mb EDO DRAM was 20 percent in 1995. This is a larger penalty than exists on the 128Mb RDRAM over the equivalent PC100 or DDR part today. Nonetheless, SDRAMs are what sells in the greatest volume today. The economics of the DRAM industry drive the volume runner to reach a price that is lower than anything else, no matter what the initial penalty was. This is why an EDO DRAM today commands a much higher price than its SDRAM counterpart. This will happen with RDRAM should it be driven to popularity. It would be quite a surprise if the 256Mb RDRAM were not cheaper than the 256Mb DDR or PC133 part over the lifetime of the density, assuming that Intel continues to increase processor clock speeds. [Bilow: Here it is a year later and RDRAM is still 4x the price of PC133. Some prediction.] Intel reminds us at every opportunity that it continues to believe in RDRAMs and will continue to promote them in the future.
PC OEMs will always go for the cheapest solution but you need to factor in the will of the supply-side players and the technology they want to push. If you do not, then you would never have seen the introduction of FPM or EDO or SDRAM, all of which had a price premium when introduced. It is a chicken and egg thing. Typically, an interested party takes the decision to break the circle, take a short-term hit, and subsidize the ramp of a new technology (this is happening now with RDRAM).
Gartner Dataquest believes Intel should have been more decisive in unequivocally backing RDRAM. Instead, the company decided to take out an insurance policy and develop an SDRAM chipset for Pentium 4. This move sent a signal of weak commitment to the PC OEMs and DRAM manufacturers that have always been hostile to Rambus. If Intel had stuck to its story that RDRAM is the "best" technical solution, then the DRAM vendors would be making it in volume by now, prices would be coming down, the RDRAM premium would be disappearing and PC OEMs would be happy to use the technology.
The Truth About RDRAM Adoption
Forget the performance argument and the cost argument. RDRAM is Intel's preferred technology, and prices would fall rapidly if RDRAMs were in volume production. So why is RDRAM not ramping to volume production now?
The issue is that there is a massive struggle for survival going on and it is not Rambus that is under threat -- it is Intel. Think about it. Intel will be a $30 billion semiconductor vendor this year. That is a lot of revenue to give up without a fight and Intel has a fight on its hands. Look again at the key industry relationships shown in Figure 1.
Intel has strong, bilateral links with Microsoft (which it relies upon to generate new, feature-rich software), PC OEMs (its customers) and Rambus (its source of high-performance memory IP).
DRAM vendors, on the other hand, are a more difficult proposition. Sure they typically want to ship 70 percent more bits this year than last year and that relies upon a robust PC market, but they do not really care whether those bits are SDRAM bits or DDRDRAM bits or RDRAM bits. DRAM vendors are quite capable of manufacturing any DRAM technology. No, the problem for DRAM vendors is fear of losing control of the DRAM technology road map and effectively becoming foundries for Intel and Rambus.
This nervousness among DRAM vendors could explain the recent behavior of both Intel and Rambus. If Intel pushes too hard and appears to be imposing its will too forcefully on the direction that PC memory technology takes, it could face a damaging backlash that will readily be supported by many in the industry. Far better to appear more circumspect and let DDRDRAM technology die quietly in the marketplace. How could this be done? Intel and Rambus could be seen to cool their relationship so that they appear less powerful, financial incentives on the use of RDRAM could be introduced and the introduction of a DDR chipset could be delayed. Sound familiar?
So, What Is Going to Happen?
There is not room for different DRAM technologies to compete in the mainstream PC space. The economics of the DRAM market support the adoption and widespread use of only a single technology. A price premium could be justified for a high-performance DRAM produced in low volume for workstations and servers, but Intel plans to cascade Pentium 4 down through the mainstream PC even more quickly than it has done in the past with previous MPU generations. The company needs to do this to frustrate AMD's efforts to gain market share.
Could different DRAM technologies exist side by side in the mainstream PC, presumably differentiated based on PC price points? This is unlikely because the existence of multiple technologies could cause DRAM product mix issues, localized supply/demand imbalances and wild fluctuations in DRAM pricing. The DRAM market is driven hard by the economies of scale, and coexisting technologies undermine these economies. So the technology that was meant to be lower performance and cheaper (for example, DDRDRAM) could end up being more expensive than the technology that was meant to be higher performance and dearer (for example, RDRAM). This has happened with every obsolescent technology in the past, be it the next lower density, or last year's interface technology.
And the Winner Is …
The so-called "best" technology does not always win. Politics and commercial expediency will see to that. But in this case, politics and commercial expediency happen to favor Intel's choice for "best" technology. Intel cannot afford to let RDRAM fail in the mainstream PC because the future success of its microprocessor business model depends on it. Therefore, we can expect to see RDRAM as the dominant DRAM technology by 2003.
By Richard Gordon, Semiconductors Europe (Richard.Gordon@Gartner.com) and Jim Handy, Memories Worldwide (Jim.Handy@Dataquest.com) with contributions from Martin Reynolds and Kevin Knox
...
google.com
-- Carl |
