Suppliers battle Rambus bug
Jack Robertson
Intel Corp. has discovered a flaw in the Direct Rambus DRAM interface that could force some chip companies to implement an extensive hardware fix across their device-mask layers, a broad range of industry sources said last week.
Several major DRAM makers said privately that in a worst-case scenario-if they are forced to make an unscheduled revision to their mask sets-it would take three months to bring the new iteration to silicon and another three to test and qualify the chip.
The bug, called a "CMOS truncated bit," causes the memory data reading to the DRAM registers to shut down, according to messages Intel sent to chip suppliers in the past two weeks. This is because the leading edge of the clock cycle terminates the register read-just the opposite of what must happen to transfer data.
Samsung Electronics Co. Ltd., which is expected to manufacture much of the Direct RDRAM produced in the industry this year, acknowledged the bug but called it a non-issue. "Samsung's Rambus is fully functional and has been demonstrated to work in fully stable systems," said Avo Kanadjian, vice president of marketing for Samsung Semiconductor Inc., San Jose.
Spokespeople and executives for three other DRAM companies-LG Semicon Co. Ltd., NEC Corp., and Toshiba Corp.-said they've received no communique regarding a Rambus flaw within the past few weeks. Toshiba, however, said it discovered a bug two months ago, since fixed, that triggered a power spike during Direct RDRAM's power-up sequence.
Intel discovered the latest problem when testing Direct RDRAM chips with its Camino chipset, according to sources. In an interview last week, Howard High, Intel's director of corporate communications, would not confirm the existence of the bug specifically, but said such discoveries are not uncommon in a new product as advanced as Direct RDRAM.
A spokeswoman for Rambus Inc., Mountain View, Calif., said she had no knowledge of any deviant register bug in Direct RDRAM.
However, David Pulling, vice president of marketing for chipset maker Reliance Computer Corp., Santa Clara, Calif., said some DRAM suppliers told him last week that they may yet have to implement an all-layers mask redesign. Most chip makers interviewed for this article declined to comment for attribution, citing contract terms with Rambus that bar them from making critical statements.
Analyst Sherry Garber of Semico Research Corp., Phoenix, added that PC OEMs had voiced concern to her about a Direct RDRAM mask redesign. And Danny Lam, director of Fisher-Holstein, a semiconductor consultancy in Wilmington, Del., said clients have told him of the problem, which he said could be significant.
According to sources, Intel told DRAM makers that they would shortly receive revised engineering specifications from Rambus to remedy the problem.
Samsung said Intel notified it of the bug, which issues a "0" instead of a "1" during the DRAM's initialization sequence, but recommended a temporary, system-level software fix. Intel then advocated a hardware redesign during the next regularly scheduled Direct RDRAM die shrink, allowing companies to avoid the time and expense of an added mask-set revision.
"Such a communication [was made by Intel], and indicated that there was a system-level work-around in place," Kanadjian said. "When the next planned die shrink is introduced, there was a recommendation that the change to Rambus be made then."
Other memory companies surveyed last week said customers may be disinclined to accept a software bandage, and may require their suppliers to make a forced mask-layer revision to correct the problem in hardware.
The change, though relatively small, would require a redesign of logic functions on the Rambus chip and could entail an all-layers mask redesign, they said. Even the smallest revision, if it involves making new masks for each of Direct RDRAM's 16 to 20 layers, would take three months to implement in silicon, and another three to qualify, chip makers agreed.
In the event of such a delay, suppliers could still get a redesigned and qualified Rambus chip to market in the late fall, about the same time as the revised timetable set by Intel for introducing its Rambus-enabled Camino chipset.
However, Semico's Garber said the bug only creates more uncertainty in the market for Direct Rambus. Many PC companies are now negotiating with Reliance Computer and Taiwan-based independent vendor Via Technologies Inc. for supplies of alternative chipsets, she said. Rather than adopt Direct RDRAM, Garber said, chip makers may be tempted to turn to PC-133 SDRAM to fill the gap caused by the Rambus delay.
techweb.com |
