PALM SPRINGS, Calif. Intel Corp. reaffirmed its commitment to the embattled Rambus memory architecture on multiple fronts this week. At the Intel Developer Forum, it announced plans to use a dual-channel Rambus memory interface to deliver 3.2 Gbytes/second of bandwidth to its next-generation Willamette processors.
Further, in a keynote address at the forum, vice president Pat Gelsinger said Intel has "no plans to support synchronous DRAM" in the Tehama chip set that will be matched with Willamette. Both products will ship in volume late this year, targeting the performance desktop market.
Finally, Intel said the integrated Timna processor, aimed at the "value" end of the market, will ship with a Rambus memory interface on-chip. Nonetheless, Intel acknowledged a stopgap strategy that will use a memory translator ASIC to support SDRAMs on Timna platforms until the price and availability of the Rambus DRAMs approach more consumer-friendly levels.
Taken together, the moves constitute a strong endorsement of Rambus despite premiums that range from 60 percent for Dell Computer Corp. (Austin, Texas) to twice as much for "second-tier" buyers, when compared to SDRAM prices.
The Camino 820 chip set that Intel is shipping today supports a 1.6-Gbyte/s peak bandwidth. But the Tehama chip set will draw on the dual-channel approach first implemented in the 840 chip set for workstations to deliver 3.2 Gbytes/s, said Peter MacWilliams, the Intel Fellow who has spearheaded Intel's Rambus efforts.
Intel put Rambus at the center of its memory road map in December 1996. Since then, wringing out the costs and getting good yields at the 800-MHz specification have proved difficult. Nevertheless, as processors move to 1 GHz and beyond, the memory industry needs to move beyond SDRAM, Intel executives argue.
At the Developer Forum's memory road map session this week, MacWilliams told a room crowded with engineers: "Any way we cut it, the transition is from SDRAM to RDRAM."
MacWilliams and Gelsinger both showed a suite of 10 benchmarks that demonstrated relatively modest performance gains from the Camino-Rambus combination for today's mainstream applications. But as the Pentium III moves to 800 MHz and beyond, the gains grow to about 10 percent compared with the BX chip set using PC100 SDRAMs.
For the Willamette processor generation, which moves to clock speeds of 1.5 GHz and beyond, Intel is projecting that the dual-channel Rambus approach will prove its worth, delivering 10 percent to 30 percent performance improvements as measured across a suite of benchmarks.
As Rambus Inc. chief executive officer Geoff Tate asked, "Why would a user want to buy a system based on a Willamette processor and then give up 10 percent to 30 percent performance by matching that with SDRAMs?"
But SDRAM technology isn't standing still, said Bert McComas, principal analyst at InQuest (Gilbert, Ariz.). The PC266 double-date-rate (DDR) SDRAMs will come closer to keeping up with Rambus than Intel's benchmarks indicate, he said. After that comes a seamless transition to the DDR-2 technology, the specifications for which are now being finalized
By not supporting SDRAMs with Willamette, "Intel has left itself open to [Advanced Micro Devices'] Athlon coming in under Willamette and filling a huge market space" in terms of cost/performance, McComas said.
But Gelsinger, general manager of Intel's desktop division, said that "with a 1.5-GHz processor, there just isn't sufficient bandwidth with SDRAMs. For a 1.5-GHz processor, it would be like giving up 500 MHz of processor speed, and that is why we have shuffled our strategy to make RDRAM the primary memory for Willamette."
(The Willamette microarchitecture also will be used in a 32-bit CPU for servers, code-named Foster, that will operate with SDRAMs and DDR SDRAMs. The initial Foster-based system will be a PC100/200 solution.)
Linking Willamette so tightly with Rambus may be a risky strategy, but the decision signals Intel's confidence that the horror stories that grabbed headlines here at the autumn developer forum are behind it, said Joe Osha, semiconductor analyst at Merrill Lynch Securities. He acknowledged that the bandwidth Rambus delivers will allow RDRAMs to replace SDRAMs eventually.
Nonetheless, Osha asked, "Why isn't Intel offering alternatives to RDRAM for Willamette, given the 20 to 25 percent price premium that RDRAM modules are likely to command when Willamette launches?"
Intel does offer a flexible approach for the Camino-Coppermine combo shipping today. For desktops, Intel will go to the CeBIT show in Europe with a "2+2" motherboard design that may be particularly welcome by VARs in the retail channel. The 2+2 board will support either Rambus in-line memory modules (RIMMs) or dual-in-line memory modules populated with SDRAMs. It's an either/or approach that supports systems starting out with SDRAMs and then switching to RDRAMs as availability improves.
At the high end of the desktop market, MacWilliams said, the premiums commanded for the RDRAMs should be acceptable. Customers paying more for the Willamette processor and the chip set won't be surprised at premiums for high-performance memory, he said.
Timna, meanwhile, will include an integrated graphics core on a die that includes a Rambus ASIC cell. But with RDRAMs projected to remain too pricey for the value end of the market, MacWilliams said, Intel has taken a contingency approach with Timna. The Rambus signaling will move to a memory transfer hub, an ASIC on the motherboard that translates to SDRAMs, albeit at a performance hit.
Asked why Intel didn't create a Timna with an SDRAM interface on the processor, MacWilliams said that such an approach would have required too many pins on the package and that the buffers and I/O transistors would bloat the die.
Timna is aimed at the sub-$600 desktop market. The integrated L2 cache, graphics and memory controller "are the perfect feature set to minimize system cost," said Albert Yu, senior vice president of Intel's microprocessor division.
The decision to use Rambus with Timna further stemmed from granularity concerns: The packet approach used by Rambus over a 16-bit channel means that small systems, such as low-end PCs or Internet appliances, can ship with 32 Mbytes of main memory. The narrow bus allows main memory to be increased in increments of 32 Mbytes, half that of today's 128-Mbyte SDRAMs.
In a tight DRAM market where prices increase rapidly, the granularity issue will play to Rambus' favor in the Internet appliance market, said Avo Kanadjian, vice president of marketing at Rambus (Mountain View, Calif.).
MacWilliams said about 7 million 128-Mbit-equivalent RDRAMs will ship in the first quarter. Kanadjian argued that if Toshiba Corp.'s production for Sony's Playstation 2 game console is taken into account Sony has said the console will use RDRAM first-quarter production might be closer to 10 million units. Sony expects to sell one million Playstation 2s in the first two days once sales commence in Japan next month. That requires Toshiba to stockpile 2 million 128-Mbit RDRAMs of 600-MHz speeds or higher.
Samsung is shipping nearly 2 million units per month now, accounting for as much as 80 percent of total RDRAM shipments, said Sherry Garber, senior vice president of Semico Research Corp. (Phoenix). As much as 70 percent of all RDRAMs are shipped to Dell, which has the most aggressive plans for Rambus implementation across its desktop line.
One source at a major RDRAM manufacturer estimated that Dell is paying a 60 percent premium for RDRAMs over SDRAMs, which sell for $16 to $18 each at the 128-Mbit density. That would put Dell's price for a 128-Mbit RDRAM at about $26. The cost to Dell compares favorably with the $50 tags that smaller buyers must pay. Further, Dell has told its DRAM suppliers that they must plan to devote more wafers to RDRAM production and steadily reduce premiums this year and next, the source said.
Toshiba, which Garber said has bet heavily on Rambus, is temporarily tied up supplying Sony, but it plans to supply PC OEMs as well. NEC Corp.'s commercial RDRAM shipments have gone largely to Hewlett-Packard for the Vectra corporate PC.
Infineon Technologies, which is gearing up for its public offering in Germany, is expected to announce impending volume RDRAM production, and Intel expects Hyundai Microelectronics to follow this spring. Micron Technology, historically a fast follower, may start volume RDRAM production this summer.
What customers want
Sang Park, president of Hyundai Microelectronics, said Hyundai is late getting into the RDRAM market, partly because the merger of Hyundai and LG Semicon temporarily pushed LG's RDRAM effort to the back burner. With the merger accomplished, Park said, "we plan to invest in RDRAM capacity additions. Our strategy is to make the kinds of DRAMs that our customers want."
Farhad Tabrizi, vice president of strategic marketing at Hyundai, phrased it differently. With Korean companies under pressure to clean up their debt, they must look to their return on investment as a top priority. For Hyundai, that means producing those DRAMs that result in the best revenue per wafer. Hyundai has developed RDRAMs at the 128-, 144-, 256- and 288-Mbit densities and is in the latter stages of testing and validation, he said.
"The tight tolerances of the Rambus memories mean that yields are sensitive to slight process changes," Tabrizi said, adding that he expects Rambus DRAMs to account for less than 5 percent of the overall DRAM market in 2001. DDR will claim 8 percent, extended-data-out and fast-page-mode DRAMs will cling to 5 to 8 percent, and SDRAMs running at 100 or 133 MHz will claim the bulk of the market next year, at 82 percent, he said.
Garber of Semico Research also remains conservative on Rambus memories, forecasting that only 80 million RDRAMs will ship this year, out of total DRAM unit production of some 5 billion units.
"This is the year that Rambus ramps, and I do foresee pretty sizable growth," Garber said. "But the DRAM makers will be capacity-limited in 2001, and I do see Rambus remaining as a niche product in the high end of the desktop market. SDRAM will continue to have a very large portion for a long time."
The Semiconductor Research group of International Data Corp. is predicting that about 11 percent of this year's DRAM shipments will be RDRAMs, and Samsung is forecasting that by the end of the year, 20 percent of shipped bits will be in RDRAMs.
MacWilliams said a tight capacity situation could help Intel's RDRAM program if SDRAM prices go up, thereby reducing the price difference between SDRAMs and RDRAMs. But he said a more pessimistic scenario also might play out: If DRAM makers are making good profits on SDRAMs, they may be slow to switch wafers to the less-proven RDRAMs, where yields are less predictable.