Re: If you have to ask, then you know NOTHING about servers...
Maybe not, but lets look at the possible loads on 1 DDR channel, just for fun.
DDR 266 bursts pages at 266 * 64 = 17 gigahertz as a serial data stream or 266 * 8 = 2.1 gigbytes per second.
The first use that comes to mind (granted, to one who knows NOTHING about servers) is pumping web pages out of DRAM if your server has a big pipe.
What can a big pipe handle?
Well, the current backbone of the internet is made up of DS-3s, OC-3s, and some OC-12s. A few OC-48s are starting to show up, and in the time frame we are looking at (a year from now), OC-96 should be on line between some some of the primary NAP (Network Access Points). So lets say your server is one hop away from the West Coast NAP, with an OC-96 connection to it, and all of your content is in RAM - how fast should your memory be?
Well, an OC-96 connection will have a theoretical speed of 51.833 * 96 = 4,976 mbits/sec or about 5MHZ as a data rate. TCP/IP over the internet won't be 100% "efficient", but I'll give you the benefit of the doubt on that one. Now returning to the original 17MHZ serial data rate that a single DDR DIMM can burst at, it should be capable of saturating the theoretical maximum capacity of the largest Internet link that will be up during the next year even if it runs at less than 30% of its burst rate.
Obviously, you will need an AMD LDT class of system bus to connect that DIMM to a set of NICs, but the bottleneck will be the system bus, not the DRAM bus.
A good site for Internet backbone info is here:
boardwatch.internet.com
How about databases? If you have a serious application, you'll have a large fiber channel disk array pumping data at your memory bus. How fast would that be? Ciprico has a nice unit:
fibrechannel.com
that will burst data at up to 100Mbytes/second - and stream it at 90. But looking up at our initial calculations, we see that our one DIMM supports bursts at 21 times that speed. Three such arrays on a system could bring the data load up to 300Mbytes/second, one seventh the capacity of a single DIMM. If the "efficiency" of DDR is only 30%, we are still OK.
If the system is running a lot of substantial CGI, ASP scripts, etc., it will be running many concurrent processes, and be heavily accessing memory as it executes the various tasks and swaps them in and out. But the accesses will be primarily short burst accesses where latency is very important. I would think that streaming large blocks of data for an extended period (where rambus could shine) probably won't be a common occurrence.
Dan |
