Caminogate: the plot thickens like caramel
Fear, uncertainty, doubt amounts to Intel blundering thud
Sources close to Intel's technologies have taken time to sift through the i820 debacle (Caminogate) and have pointed to fundamental flaws in the chip firm's validation plans.
The reliable source points to an Intel developer page dealing with RDRAM validation and says that the validation process is statistically meaningless, given the number of components (from two to five) that were tested. This is the page and this the results page he has in mind.
The sample size, he notes is between one and four. "I would certainly have confidence in a telephone survey that called one to four people out of a target demographic...not!", he said. The i820 and the i840, the latter of which was formally released today, use identical Rambus interface circuitry, he claimed.
And there are other mobos rather than Dell (sorry Intel's) that fail with three RIMM modules. "These failures have to do with multiple back to back pipelined transfers from multiple devices on the bus. The guys at Hard OCP were given accurate information. This is a key feature of a Rambus system, to have multiple devices accessed simultaneously and driving the bus in a pipelined back to back fashion. This is where the only performance advantage for the bus will come from," he says.
And the failures are not just poor board design, with some serious theoretical issues involved -- resonance being the key one. "Buses are running into quarter and half wavelength resonance effects which cause voltage margin and timing margin violations. No re-engineering of the drivers, board, or receivers will fix this problem with three RIMM systems.
"Resonance effects are not limited to three RIMM systems, however, the effect here is much greater. There are reasons to believe that resonance problems can occur with two RIMM systems under extreme circumstances. It would literally take years of simulation time to simulate most of the possiblities," he adds.
However, poorly designed boards will exhibit the problems more frequently than better designed boards, while server and workstation mobos (i840) are designed with more margin, and hence less chance of failure. He makes the somewhat interesting point that Caramel (i840) systems were not designed intelligently with the foresight that three RIMM busses would cause trouble.
Instead, the i840 systems were designed with dual Rambus channels with two RIMMs per channel, and bus extenders when more modules were desired for greater size. Dual channels with two RIMMs each (four in total) is about all you can put in a standard form factor PC. "So the workstation chip boys lucked out," he says.
Neverthless, he claims that i840 bus extender and memory riser technology is not as good as it might be. He reminds The Register of a story we wrote at IDF, pointing out that DDR, not Rambus, is the future for servers and for the Itanium Merced platform. And he also claims that Alfredo Moncayo, Physical Layer Architect for Rambus has left Intel. Moncayo invented the Rambus channel.
This week, in the US, a paper on a 1.6Gb/s memory bus is being presented at a conference, in which Moncayo and Intel Rambus designer Michael Leddige will speak. According to the source, the mobo group that developed the Vancouver listened to technologists in Intel's platform architecture group, and so produced a design that didn't work. Mobo groups have to develop to the specs that PAL decides.
Down at Intel Folsom, the chipset group implements the chipset design and creates the OEM design rules, so sitting in the middle of the minefield, while customers, including Dell and Compaq, with them. So while most RDRAM chips function, the i820 functions, the RIMMs function, and the systems were designed according to the right spec, the systems nevertheless fail because of fundamental problems in physics. Oh, and the 133MHz DRAM chipset interfaces do not meet the right timing in systems either. This, of course is architecture. And as we keep pointing out, that can differ from marchitecture...
And talking of marchitecture... The dead hand of Chipzilla’s marketing goons appears to have fallen on the shoulder of the plucky US website taking early orders for Camino mobos. US outlet NECX Direct had full details of the Cape Cod and Vancouver mobos up on its site last week (see story Camino mobos go on sale), but the delayed (aka dysfunctional) Intel parts are now conspicuous by their absence. This calls into question Great Stan VP Paul Otellini’s somewhat rash claim that the i820 will ship before the end of the year.®