I dont know if this article has been posted.Found it on Rambus site. ednmag.com ultracompetitive graphics business.
Brian Dipert, Technical Editor
The press has spilled a lot of ink over the last few years regarding Intel's activities in graphics and memory. With the exception of a late-1997 cover story on DRAM ("Advanced DRAM puts you in the fast lane," EDN, Oct 9, 1997, pg 52) and a few pieces in EDN's Leading Edge, I've avoided adding my voice to the din. In the absence of official notification from Intel and its partners, most of what you hear is only speculation—rumors that newshounds print one week and contradict the next. But, after enduring even more debate and gossip at the Platform 99 conference (San Jose, CA) a few weeks back, I've decided to say a few words.
In 1996, Intel made three significant acquisition and partnership announcements. First, the company bought Chips and Technologies, for its low-power 2-D graphics-core expertise. Then, it announced a partnership with Real3D to develop a 3-D graphics core. (The two cores combined to form Intel's i740.) Finally, it announced a partnership with and investment in Rambus Inc to develop a next-generation DRAM architecture that it would derive from Base and Concurrent Rambus DRAMs (RDRAM).
Few analysts and editors could figure out why Intel wanted to get into the low-margin, ultracompetitive graphics business. Vague explanations from Intel that other vendors weren't advancing the technology (and therefore the PC platform), boosting performance, adding features, and lowering costs fast enough, weren't supported by data and didn't fool anyone. For the last several years, mainstream PC graphics hardware capabilities have far exceeded the needs of the operating systems, application-programming interfaces, and applications that call them.
Most of the accusations regarding Intel's DRAM activities concerned control: increasing microprocessor and core-logic market dominance via early access to the Rambus technology, a hefty return on their investment thanks to royalty-driven Rambus stock-price appreciation, and perhaps an eventual acquisition of Rambus. Intel reiterated that system performance was behind the Rambus activities. Simple evolutionary DRAM architecture improvements would soon run out of headroom, and because a change was inevitable, 1996 was as good a time as any to start the process.
Three subsequent years of DRAM unprofitability, coupled with the emergence of high-performance double-data-rate synchronous DRAM (DDR SDRAM), have diminished but not destroyed the viability of Intel's claims. But there's one angle Intel hasn't talked about, perhaps because the company has seen the legal pain that the word "bundling" is causing partner Microsoft, and because the Federal Trade Commission has its eyes on Intel, too. I think this angle is the most compelling reason for Intel's graphics and memory actions.
Consider a single chip containing a CPU, core logic, graphics, and some amount of cache. Intel's dabbled here before, with its 386SL and 486SL. National Semiconductor's recent Geode announcements prove that the vision is technically feasible, even with today's process technologies. (Economic viability is undetermined.) Intel has already integrated graphics into the core logic with their i810, as have SiS, Via and others. Integration, targeting low-end PCs, will only become more compelling as sub-0.2-µm lithographies ramp into production, and reducing pin count will minimize cost. Such a chip's die size might not be bond-pad limited today, especially if it contains L2 cache, but at 0.15 µm or so, it [ensp]could be.
A 76-pin (including ground traces), 400-MHz (800-Mbps/pin) Rambus channel, at roughly the same peak bandwidth as 160-pin, 100-MHz (200-Mbps/pin) DDR SDRAM, requires less than half of the DDR SDRAM interface pin count. Intel claims that coupling its CPUs' multiple-transaction buses to RDRAM's deeply pipelined architecture results in a much higher effective bandwidth than that of the DDR SDRAM alternative. If you further integrate the graphics on the CPU, you can also eliminate the approximately 70-pin accelerated graphics port (AGP) bus. I know it's hard for power-users to imagine a PC with a graphics subsystem that you cannot upgrade, and I agree that Silicon Graphics made a mistake here with its Windows NT line, but remember, we're talking about low-cost PCs, not high-end workstations.
Buyers of low-end systems don't want to replace boards and software drivers. They don't read computer and gaming magazines to analyze the benchmarks. They probably won't run high-end applications that heavily stress graphics. (Even I'm still happily using a two-year-old graphics board!). Low-end PCs come with small monitors that don't require the resolutions and color depths that might bring entry-level graphics accelerators to their knees. And at the price of these systems, the average consumer will replace the PC (and its graphics) every few years or so, anyway.
To those of you who grumble about integration and about leveraging success in one area to increase probability of success in another, I say, "Get over it." Intel doesn't have exclusive access to the benefits of Moore's Law just because the guy who came up with the law works there. Via has figured out where the market's going and, through partnerships and acquisitions, is headed in the same direction. AMD, with Altheon, has shown that it has core-logic expertise to match its CPU prowess. And Rise Technology, or any other not-yet-public x86 vendor, can also go down the same integration road.
I shake my head every time I see an analyst predicting that Direct RDRAM (DRDRAM) will never penetrate the low end of the PC market. With all its money, Intel has the luxury of investing for the long-term. Anyone who restricts DRDRAM to a high-end niche needs to refocus his binoculars or polish his crystal ball. After years of development effort by thousands of engineers spending millions if not billions of dollars, I struggle to imagine a scenario in which DRDRAM will be unsuccessful. Non-PC platforms, such as Sony's upcoming Playstation 2, will only increase the probability of DRDRAM's acceptance.
Most of the companies criticizing Rambus are behind schedule in their own DRDRAM programs or are understating their yields, to deceive their competition or extract the maximum profit margin from their customers. Yield, which translates to cost, will fix itself in a process generation or two or with a few design tweaks. You heard the same grumbling during the EDO-to-SDRAM transition years ago. Similarly, many critics of integrated graphics have just not yet secured their own partnerships with core-logic and CPU vendors.
Be careful when you hear someone bad-mouthing a new technology. Dig a little to find out what vested interests they might have in ensuring that this approach doesn't succeed. Disparaging Intel, Microsoft, and Rambus is very fashionable these days, and many people in the editorial community are taking the easy way out by doing it.
Sorry, I refuse to follow this trend. If I'm wrong, that's fine, but I'm not going to simply regurgitate what is politically correct. Then again, I have never been very fashionable. Just ask my wife, or you can check out my wardrobe.
--------------------------------------------------------------------------------
Author Information
Contact Technical Editor Brian Dipert at bdipert@pacbell.net. |
