To: dougSF30 who wrote (198045 ) 5/19/2006 9:07:18 PM From: pgerassi Read Replies (1) | Respond to of 275872 Doug: These numbers were pulled from other sources. I don't see any links to sources that say otherwise (datasheets preferred). Of course you don't want to hear that 8.5W gets used by the MTH in Blackford's chipset group for each FB-DIMM channel, but what did you expect? That the FB-DIMM controller uses no power at all? Get real. The MTHs in a typical P4 chipset uses power to control each DDR2 channel whether they are in a single die or multiple ones. Its just that DDR/2/3 controllers use a lot less power, less than 5W for both DDR channels in a K8. And Doug, FB-DIMMs are serially connected. To get to the last one, you have to go through each one in the chain. That is the same problem as Rambus RDIMMs had. So all of the AMBs must be active in a channel. As to all of them not being active, get real on that too. When running on either Linux or any Windows based server, any unallocated memory is being used to cache the HDs. That means that all of the DRAM is active in a running active server. Only when idle, will the DRAMs be mostly inactive. But that means no work is getting done. Virtualization is a means to get more processing done with a fixed amount of hardware. So you are replacing x mostly idle servers with a few active servers. Thus the idle time will be small, if nonexistant. Even if one of the four cores in a 2S Woodcrest server is active and the rest idle, all of the memory will be active. Prefetching increases the amount being active. All in all, those FB-DIMMs will be active and being used, almost all of the time. The power consumed is worse than with a 2S Opteron server, all else being equal (82W vs 10W not including either DRAM dies or CPU cores for the 4 DIMM case, 184W vs 10W for the 16 DIMM case). That is the tradeoff Intel made in keeping external NBs, MTHs and FB-DIMMs instead of ODMC, ODNB and DDR2. Pete