Does the 1066MHz FSB Improve Memory Performance?

Quite possibly the biggest feature of the 1066MHz FSB today is the fact that it runs at a clock multiple of DDR2-533's frequency. Why is that such a big feature? It's analogy time:

If two people are having a conversation and they can both talk and listen at the same rate, then the conversation will flow as smoothly as possible. If person A talks and listens slower than person B, then person B will always be waiting for person A instead of communicating as fast as possible - a frustrating situation for those that have been here before. Running a FSB and memory bus asynchronously is just as frustrating to the CPU; if the two frequencies aren't synchronous then there is an additional latency penalty incurred while transferring data between the two buses and because of that additional latency penalty, there is a reduction in usable bandwidth.

With the original 925X chipset we were a bit unhappy to see that the Pentium 4's 800MHz FSB was paired with DDR2-533, creating one of those frustrating asynchronous situations. But with a 1066MHz FSB (266MHz x 4), the 925XE can communicate synchronously with DDR2-533 (266MHz x 2), thus reducing memory latency and increasing memory bandwidth in theory. What do we see in practice? To answer this question we look to two trusted measures of memory bandwidth and latency: CacheMem and ScienceMark 2.0.

First looking at latency we see that with the 1066MHz FSB, memory latency with DDR2-533 looks like it's hardly improved. Running the FSB at 1066MHz manages to shave off a just a few clock cycles.

DDR2-533 3-3-3-12 Latency Comparison - Cachemem
 
1066MHz FSB
800MHz FSB
Performance Improvement
512-byte stride - 32MB block
226 cycles
227 cycles
0.4%
1k stride - 32MB block
239 cycles
241 cycles
0.8%
2k stride - 32MB block
266 cycles
266 cycles
0%
4k stride - 32MB block
311 cycles
311 cycles
0%

We look at ScienceMark and see the same basic situation but with slightly improved performance; looking at the absolute latency values in nanoseconds we see that the 1066MHz FSB manages to reduce memory latency by around 2 - 6%.

DDR2-533 3-3-3-12 Latency Comparison - ScienceMark 2.0
 
1066MHz FSB
800MHz FSB
Performance Improvement
16-byte stride
4.69 ns
5 ns
6.2%
64-byte stride
17.5 ns
18.12 ns
3.4%
256-byte stride
73.43 ns
75.93 ns
3.3%
512-byte stride
75.93 ns
77.81 ns
2.4%

The reduction in latency isn't insignificant under ScienceMark, but what about its affects on memory bandwidth?

Looking at Cachemem once again we see an increase in memory bandwidth of just under 4%.

DDR2-533 3-3-3-12 Bandwidth Comparison
 
1066MHz FSB
800MHz FSB
Performance Improvement
ScienceMark 2.0
4742.02 MB/s
4347.63 MB/s
9.1%
CacheMem
3455.3 MB/s
3324.7 MB/s
3.9%

ScienceMark appears more considerate of Intel's hard work and shows a 9% increase in memory bandwidth. The thing to keep in mind here is that the improvement in memory bandwidth will depend on the types of applications run, but the ScienceMark and Cachemem results should give you an indication of the range of improvements to be expected in applications that are memory bandwidth limited. In those applications that aren't currently bound by memory bandwidth, the impact will be much less.

What is important to keep in mind here is that DDR2 is still running at relatively high latencies. Even while running synchronously with the FSB, the 925XE and DDR2 combination still posts higher memory access latencies than 875/DDR400 platforms thanks to DDR2's high memory timings. We were able to run 3-3-3-12 timings on our DDR2 test platform by keeping memory voltage at a safe, but overclocked, 2.0V, but anything faster than that was too much for today's DDR2-533. It will take even lower latency DDR2 in order for even the 925XE platform to show some further performance advantages.

Given the relatively small increases in memory bandwidth and decreases in latency, the extra 66MHz of the 3.46EE will have to go a long way in order to gain any more ground for Intel. Let's see how things shape up in some real world tests.

Index Does it Improve Real World Performance?
Comments Locked

63 Comments

View All Comments

  • Beenthere - Sunday, October 31, 2004 - link

    Intel has simply run out of Hail Mary solutions to their unending design, engineering, production, sales, management, and marketing problems. Even Wall Street knows this by now.
  • Wesley Fink - Sunday, October 31, 2004 - link

    For those who asked, the 1000 lot Intel price for the 3.46EE is $999.
  • coldpower27 - Sunday, October 31, 2004 - link

    Well you could get a direct comparison between the Athlon 64 3700+ vs the Pentium 4 560 as those 2 processors are priced pretty directly against each other on Newegg, though their MSRP differ in actuality.

    64Bit Windows isn't likely to be released until Prescott 2M with Intel EM64T is released in Q1 2005. We will have to see though if Microsoft will released in 2005 WinXP 64.
  • jimmy43 - Sunday, October 31, 2004 - link

    #14 I was thinking the same thing. The 3500 would probably still win or tie in most of the categories and it costs nearly half the price of an intel 560. I feel like Anand is trying to be fair to both companies and reccomending a bit of both. Realistically, AMD has Intel beat in every market segment... by alot. It's also funny how everyone is COMPLETELY forgetting that AMD's proccessors are 64 bit so in a year or so, you will get a considerable free speed boost and youl be able to run the latest OS. Is that not a huge advantage? Come on, people need to stop overlooking that its really bugging me.
  • DukeN - Sunday, October 31, 2004 - link

    Once again, Intel shows why it's the Sony of the CPU world with terrible products terribly overpriced.
  • Gnoad - Sunday, October 31, 2004 - link

    I might have missed it, but how much will Intel price this at? Considering it's an EE, one can guess about $900. If thats true, they MIGHT sell 3 or 4 of them.
  • skunkbuster - Sunday, October 31, 2004 - link

    underdog in terms of market share
  • GhandiInstinct - Sunday, October 31, 2004 - link

    Since when did the world spin where a chip that is superior in 90% of chip tasks is the underdog?
  • stephenbrooks - Sunday, October 31, 2004 - link

    So... I was thinking of investing some money in shares. You don't think AMD might happen to be a good bet right around now, would you?
  • SLIM - Sunday, October 31, 2004 - link

    Great review as always, but there's always room for improvement:)
    [/begin nitpicking]
    "So in the end, who takes the crown? AMD or Intel? The 3800+ took four category wins, while the Pentium 4 560 only took two, however with the exception of the gaming and professional apps category, AMD's victories were not overwhelming - especially once you take into account the fact that the 3800+ is priced much higher than the Pentium 4 560. Now that you can purchase at least a couple of 915 based motherboards for less than $130 the total cost of ownership for the Intel platform doesn't eat into the CPU price advantage. For the most part we'd say the 3800+ is faster than the Pentium 4 560 but not always worth the added cost. It's unusual but in many cases, the Pentium 4 560 is actually the bargain high-end chip of the two."

    Alrighty, two comments:
    1) It's bad science to make a detailed comparison, and then in the conclusion talk about switching the chipset and memory in order to make the price comparison hold up. (Maybe include the numbers from a 915 review to back up the assertion that the 560 will still perform just as well with 915/DDR).

    2) I'd be curious to see how the 3500+ would hold up in these same comparisons since it is about $150 cheaper than the 560.

    Bonus nitpick:
    4 of the graphs don't include the new 3.46ee (ACD on page 9 and 3 games benches); I don't know if that was intentional or not. [/end nitpicking]

    Thanks again for the best reviews.

Log in

Don't have an account? Sign up now