Performance and Scalability of the new Opteron

In the interest of fairness, I've been monitoring Microsoft's Exchange Server benchmark page. As you may recall, when the Opteron was launched, scores for 2-way and 4-way servers based on the chip were posted to this page.

As discussed here, the not yet available Opteron platform (running at 1.80GHz) defeated the fastest competing Xeon (2-way, 2.80GHz) and Xeon MP (4-way, 2.00GHz) chips by 12.2% and 17.6%, respectively.

As of this month, two new scores have been posted, one which is particularly notable. Both scores are from 2-way Xeon servers. These servers are running at 3.06GHz, and the faster of the two has a score of 11000, a jump of 12.2%. Curiously, this is exactly enough of a boost to make the fastest 2-way Xeon the same speed as the fastest 2-way Opteron. Rather than chalk it up to an unlikely coincidence, I'm tempted to speculate that there's a bit of rounding involved in the score reporting process. No matter, though; the basic idea is that the Xeon and the Opteron are currently evenly matched in 2-way configurations.

It is important to point out that the Opteron system is working with twice as much level 2 cache memory as the Xeon, whereas the Xeon are working with a 70% frequency advantage. There are also differences in the disk subsystem, and that can count for a lot (for example, the other new Xeon score is similarly configured in terms of frequency and cache memory but scores 13.6% lower even with a slightly newer version of Windows 2000 Advanced Server).

The 4-way Opteron still maintains its unchallenged lead of 17.6% above the closest competing 4-way server. But there's always next month. ;)

Microsoft is oomphing around about the Opteron. One of the processor's strengths is apparently n-way performance in heavy server situations. A 1.8GHz 4-way Opteron runs about 17.6% faster than the highest listed 4-way Xeon MP. If the performance scales more or less linearly with frequency, that performance level might be met by a 4-way Xeon MP clocked at 2.6GHz or higher, but the Xeon MP is currently sold at frequencies no higher than 2.0GHz. Unlike the non-MP Xeon, which has a maximum of 512KB L2 cache and no L3 cache, the Xeon MP can be equipped with up to 512KB L2 and and up to 2MB L3 cache. Additionally, the Xeon is limited to single or dual processor systems, whereas the Xeon MP can be used in "feature-rich 4-way, 8-way, and even greater configurations". The reason why this processor runs at a much lower frequency is that systems designed for higher level multiprocessing typically require more stringent manufacture-level testing and must pass higher standards for operation. Whereas a regular consumer can live with their processor, for example, overheating and resetting once or twice in its lifetime, a professional user has to be guaranteed something more closely approximating perfection (mind you, I'm not trying to imply that the regular Xeons are prone to overheating -- I'm just trying to present something analogous, something that'll fit in with the difference between consumer and corporate product standards).
Anyway, I believe that Intel will be able to ramp the Xeon MP frequency higher than their roadmap currently indicates if they feel that they need to. Keep in mind that AMD's penetration in the server market is relatively low (heck, a single percentage point gets more respect than AMD's server offerings! so Intel could easily continue to dominate in this market based on momentum and reputation alone without any serious acceleration of their n-way x86 schedule. Still, the 2.0GHz Xeon MP (with 2MB L3) was, as of last November, selling at a price of $3692ea in 1k packs. The price is probably lower now, but I do suspect that the current price is still substantially higher than that of the new Opterons, so we may see some strong downward price movement on Intel's part sooner or later.
Still, the Opteron seems to go well against the highly clocked vanilla Xeon in this particular benchmark. The 2-way 1.8GHz Opteron configuration is listed with a score 12.2% higher than the fastest listed 2-way Xeon system, in this case an IBM eServer xSeries 235 with two 2.8GHz Xeons. Two things to note about both the 2-way and 4-way comparisons: Firstly, Intel recently introduced an 800MHz platform that connects to two channels of DDR memory at an insanely jiffy 6.4GB/s.
Secondly, the AMD scores were posted this month, while most of the Intel scores were posted months ago, so it'd be only fair to see updates on the Intel side, as we'd likely see this performance gap narrow with newer configurations.
On the flip side, the AMD platform was penalized by the nature of the operating system. Obviously, since Microsoft's x86-64 version of Windows NT is not yet ready for the market, the tests were done with the 32-bit Windows 2000 Server operating system. While performance boosts are not guaranteed with the upcoming Opteron-targeted operating system, it is not unrealistic to conject that full support for the higher memory addressability (native access to larger than 4GB memory spaces is probably better than using clever but probably less efficient methods of memory access such as PAE) as well as better handling of the K8 platform's NUMA would result in a nontrivial performance boost.