Intel Pushes Back with Xeon 5100
conq writes "BusinessWeek has a piece on Intel's newest chip, the Xeon 5100, which many consider might be the chip that will llow them to stop losing ground to AMD. From the article: 'During the presentation, Intel ran the now-standard comparison test against AMD's highest performing chip, handily beating the system in a speed test. And in a jab at AMD execs, who handed kill-o-watt meters to analysts at the outfit's recent technology day, Intel execs used the same device to measure the new Xeon 5100 system's performance — gauged to be 7 watts better than that of the AMD-based system.'"
Intel has by far the largest fabrication capacity of any chipmaker in the world. Both IBM's and AMD's fab capacities are much lower (AMD has used IBM's fab to help meet demand). IBM's inability to produce high numbers and high yields led to the Intel switch. Remember the delay in introducing the iMac G5? Apple had the design ready, IBM couldn't produce the chips. Result: months go by without any iMacs to sell. More than anything technical reason, IBM was bad for Apple's bottom line.
Which is why these 64-bit Linux benchmarks show that Woodcrest scales as good as (and sometimes better than) Opterons at 4p. The vast majority of x86 servers are in the 4p range. Even Opterons have a worse-than-expected scaling issue past 4p, anyway, if you bother to look around to find the benchmarks.
The Optron's scaling issues beyond 4P is not "worse then expected," because it is entirely expected of the architecture.
The high-end Opteron has 3 HT links. This means it can work with up to 8 sockets "gluelessly," but it really performans much better with 4-socket systems. The architecture for a 4-way Opteron server uses the extra HT link to reduice the number of hops, so only one case has two hops.
But you can imagine that the 8-way configurations have a much higher average number of hops between processors, PLUS much more data flowing over the same HT links. No, the K8 Opteron is not really designed well for 8-socket systems.
But K8L IS designed for 8-socket systems.
Take a look at a page on this in the K8L preview article on Real World Technologies. Adding a 4th HT link will really make a difference.
4-socket K8L systems benefit because they take advantage of the 4 HT links to provide 1-hop latency to all sockets in the mesh, and can now have external I/O hooked up to ALL processors.
8-socket K8L systems take advantage of two things: the extra HT link is beneficial, and the advanced mesh created by splitting up the HT bus widths means MUCH better performance for 8-way systems.
Woodcrest is impressive as hell, but I will tell you one thing: there's no way in hell it's going to scale well beyond 4-socket systems. This is for the same reasons that have been holding back performance on 4-way Xeon syetems (reduced bus speeds with 4 processors on the bus, too much traffic). The Dual-Independent Bus allows Intel to scale well to 4-way, but no higher. K8L will allow for glueless scaling to 8-way, and will still provide a a cheaper solution than Intel's Dual-Independent Bus for 4-way chipsets and motherboard designs.
Man is the animal that laughs.
And occasionally whores for Karma.