TGPTNDR, Re: "We haven't had an X-86 server that wasn't maxed out on conventional ram for 5 years. In server land, IMO, Ram is lots more important than CPU sockets."
Which is a limitation of the chipset, not the CPUs. Any 32-bit Pentium based processor can address up to 64GB of memory. It's up to the chipset how much will be supported. Since it's hard to find more than 1GB on a single DIMM (unless you use very dense double-stacked DIMMs, which isn't supported by every chipset), the amount of supported memory has to do with the number of memory slots that can be handled electrically by the chipset. For entry servers, it is usually 4 slots, so you won't find any entry level servers - 64-bit or 32-bit - that support more than 4GB of memory. For mid-range servers, the chipsets usually employ memory repeaters, which can double, triple, or sometimes quadruple the number of available slots, allowing for up to 16GB of memory. High end servers usually have multiple internal CPU or bus nodes, each with memory interfaces and repeaters, so only when investing in the higher end systems will you find capacities reaching 64GB these days. The Xeon can handle 64GB of memory, making it fully capable of addressing the largest of memory configurations. Only in mainframes and similar very high end enterprise or HPC systems do you see memory capacities exceeding 64GB, and these cost millions to build and deploy.
However, I will agree that the future will require additional memory capacities. Every 18 months, the density of memory doubles. I expect 2GB DIMMs to be available next year. By late 2004, we might see 4GB DIMMs. By 2006, 8GB DIMMs. Around this time, I'd expect the memory capacity for entry level servers to once again be roughly 4 times the largest commercial memory density; in this case, 4*8=32GB. Mid-range servers using memory repeaters at added cost will be able to address 128GB. High end servers will probably max out around 512GB. By this time, Intel will need more than a 36-bit addressable range for the Xeon, unless they want to be relegated to entry level.
However, I believe it's a fallacy to assume that 32-bit servers will be eliminated by this road block. After all, if the Xeon can address 36-bits of memory, what's to stop Intel from putting 40-bits or 44-bits into their next generation Xeon product? Or even their consumer Pentium chips? Granted that flat memory spaces are preferable from a performance standpoint, but if the overhead for segmented memory is small relative to the total performance of the next generation systems, then it should become a moot point.
What should concern you more is the inability of the Hammer to address more than 4GB unless it is running a 64-bit OS, the timing of which is still questionable. Linux looks to be ready, but 64-bit Windows timing may not be until Longhorn in late 2004, and even then, 64-bit apps don't come overnight. I see 32-bit Hammer running into a memory wall before the Xeon. Given my projection for memory requirements, 32-bit Hammer in 2004 will barely have the memory requirements for entry level. 64-bit Hammer will probably begin ramping in 2005, or at least that's my expectation, given the heavy lifting required for 64-bit enablement. AMD would have been better off simply adding a segmented address range to their current parts.
wbmw |
