Preliminary GMA X3000 benchmarks, and thoughts on Conroe vs. Woodcrest

Discussion in 'Mac Pro' started by DavidC1, Aug 11, 2006.

  1. DavidC1 macrumors member

    Joined:
    Jul 26, 2006
    #1
    http://www.pconline.com.cn/market/sh/shoppingguide/changshang/0608/844892.html

    Here you can see with same E6300 CPU GMA950 is substantially FASTER than the GMA X3000. Of course with Quake 3 scores having THAT big of a difference, the lower scores probably can be attributed much to the drivers.

    Conroe vs. Woodcrest

    FYI, if it wasn't for dual socket capability, having Core 2 would have been faster than Woodcrest despite Woodcrest having quad-channel RAM and 25% faster FSB clock. The high latency of FB-DIMM+lower performing memory controller on the Woodcrest system makes Woodcrest system SLOWER than Core 2 based systems with 965/975 chipsets. The overall latency reported is TWICE as higher on the Woodcrest for memory than Core 2, for those thinking Woodcrest is faster because of the 25% higher FSB...
     
  2. MacsAttack macrumors 6502a

    Joined:
    Jul 2, 2006
    Location:
    Scotland
    #2
    "Faster" is such a relative term.

    Latency is only half the story. Lots of heavy duty number crunching (the sort of thing you do on a workstation for example) is more dependent on bandwidth. Which the Woodcrest memory layout provides in spades.

    Weird coincidence that. Wot?
     
  3. DavidC1 thread starter macrumors member

    Joined:
    Jul 26, 2006
    #3
    Not really. Theory is not reality. Look here: http://www.tomshardware.com/2006/06...10.html#dual_vs_quad_channel_memory_bandwidth

    http://www.gamepc.com/labs/view_content.asp?id=woodcrest&page=5&cookie_test=1

    THG: Xeon Woodcrest:
    Quad channel FB DDR2-533: 5732MB/sec 5744MB/sec(Int/FP)
    Dual channel FB DDR2-533: 3807MB/sec(Int/FP memory bandwidth identical)

    GamePC: 6GB/sec with Quad channel FB DDR2-667 memory

    http://www.tomshardware.com/2006/08/02/thg_tuning_test/page15.html

    http://www.firingsquad.com/hardware/core_2_memory_tuning/page12.asp
    THG: Core 2 Extreme:
    Dual channel DDR2-800 5874MB/sec 5860MB/sec(Int/FP)

    FiringSquad:
    5416/5431MB/sec(Int/FP) using Dual channel CL5 DDR2-667
    5735/5744MB/sec(Int/FP) using Dual channel CL5 DDR2-800

    Did I mention you can use even faster DDR2 memory for Core 2??

    Less than 6% advantage in bandwidth and TWICE the latency.

    No wonder SpecFP on Woodcrest is outperformed by Core 2 Extreme clocked lower, and SpecFP is also representative of workstation.
     
  4. Fadl macrumors member

    Joined:
    Aug 7, 2001
    #4
    Memory bandwith isn't that importing for the new Core architecture.
    DDR2-800 is only 1-2% faster than DDR2-667. More important is the FSB Speed(when running 1:1)
     
  5. Silentwave macrumors 68000

    Joined:
    May 26, 2006
    Location:
    Gainesville, FL
    #5
    And what about the other advantages of FB-DIMM?
     
  6. LoveMacMini macrumors regular

    Joined:
    Jul 30, 2006
    #6
    the advantage of fb dimm won't show themselves for another year or two

    it's expensive right now since the market demand is slim, and the heat output is intense thanks to those heatsinks.

    personally i'm going to shut up until DDR3 comes out
     
  7. YoYoMa macrumors 6502

    Joined:
    Aug 3, 2006
    #7
    Why is DDR2-800 only 1 or 2 percent faster than DDR2-667???
     
  8. YoYoMa macrumors 6502

    Joined:
    Aug 3, 2006
    #8
    Sorry, but also, what advantages can we see FB-RAM gain in the next year or two?
     
  9. Eidorian macrumors Penryn

    Eidorian

    Joined:
    Mar 23, 2005
    Location:
    Indianapolis
    #9
    Faster bus but increased latency. From what I've read on the X3000's drivers, they're terrible right now.
     
  10. YoYoMa macrumors 6502

    Joined:
    Aug 3, 2006
    #10
    Is this also going to be the case when MBP/MB goes Santa Rosa and gains the 800mhz bus? Just a 1 to 2 percent increase?
     
  11. Counterfit macrumors G3

    Counterfit

    Joined:
    Aug 20, 2003
    Location:
    sitting on your shoulder
    #11
    Heatsinks don't make more heat.
     
  12. LoveMacMini macrumors regular

    Joined:
    Jul 30, 2006
    #12
    whoops, don't drink and type kids

    I meant the heat is so great they need heatsinks! thanks for the correction :p
     
  13. Eidorian macrumors Penryn

    Eidorian

    Joined:
    Mar 23, 2005
    Location:
    Indianapolis
    #13
    It's more around the other features of the new Merom platform. Like 802.11n, Robson, and the X3000.
     
  14. LoveMacMini macrumors regular

    Joined:
    Jul 30, 2006
    #14
    ding ding ding
     
  15. Eidorian macrumors Penryn

    Eidorian

    Joined:
    Mar 23, 2005
    Location:
    Indianapolis
    #15
  16. YoYoMa macrumors 6502

    Joined:
    Aug 3, 2006
    #16
    Hmmmm, so it will be the case for Santa Rosa as well then? I had no idea..
     
  17. DavidC1 thread starter macrumors member

    Joined:
    Jul 26, 2006
    #17
    Well you gotta remember 1-2% performance difference between DDR2-667 and DDR2-800 is with Core 2 using same 1066MHz FSB. What gain will be there after since DDR2-667 exceeds the FSB bandwidth. But Santa Rosa's 800MHz FSB comes along with support with DDR2-800 so the gain will be much greater. Perhaps 3-5%.

    Did you expect 10% to be important then?? When you are saying its not important what are you comparing to?? DDR2-800 isn't that faster because its on the same 1066MHz FSB and DDR2-667 is enough.

    Go back to Pentium III days and you'll see 33% improvement in FSB and RAM resulted in 4-6% performance difference, and almost none when RAM speeds went above FSB speeds.

    You'll see that the gains for so-called bandwidth hungry Pentium 4 is also negligible: http://www.anandtech.com/showdoc.aspx?i=1615&p=6

    Did people forget about the days when 5% was the performance increase with new CPUs compared to the previous top end??

    My point was against MacAttack. Xeons are slower than Core 2.
     
  18. DavidC1 thread starter macrumors member

    Joined:
    Jul 26, 2006
    #18
    Well the only CPUs to use FB-DIMMs now are Xeon based on the Core architecture.
     
  19. Fadl macrumors member

    Joined:
    Aug 7, 2001
    #19
    Xeons are faster cause of their higher FSB (1333 vs. 1066)
     
  20. DavidC1 thread starter macrumors member

    Joined:
    Jul 26, 2006
    #20
    Sigh. Do you make a comment disregarding the previous ones?? As I said it again, the poorer chipset performance+FB-DIMM makes Xeon SLOWER than Core 2. I don't want to be harsh but you sound very ignorant.

    10% higher bandwidth
    2x the latency

    over Core 2

    achieved by
    2x DIMM channels
    25% higher FSB

    = slower

    SpecCPU benchmarks support the very fact. Xeon 5160 with 2% higher clock speed and 25% faster FSB clock is outperformed by Core 2 Extreme in memory subsystem intensive SpecFP by 10%!!!
     
  21. Fadl macrumors member

    Joined:
    Aug 7, 2001
    #21
    The latency of the Ram shouldn't affect the performance that much.
    We have to bench similar systems with core 2 duo and xeon against to be sure.
    Why don't you post some benchmarks?
     
  22. DavidC1 thread starter macrumors member

    Joined:
    Jul 26, 2006
    #22
    It's not the latency of the RAM only. It's the latency of the whole memory subsystem. You can obviously see that CL5 and CL3 comparisons in memory latency tests do not reach 67%, it reaches only 13-15%(http://www.xbitlabs.com/articles/cpu/display/amd-socket-am2_11.html)
    Chipsets for PCs are always more configured for higher performance compared to the workstation/server chipsets which are made for better stability/flexibility.

    When you see 2x the higher latency though, there's something wrong. The pattern has been that Xeons are slower per clock than the desktop equivalents since the chipsets have lower performance. One of the biggest reasons that A64 is faster than AXP is the lower latency achieved by the integrated memory controller.

    So far here's the closest I can get:
    Xeon 5160: http://pc.watch.impress.co.jp/docs/2006/0801/tawada81.htm
    Core 2: http://pc.watch.impress.co.jp/docs/2006/0714/tawada79.htm

    Hard to conclude on the other benches though, as most of the apps are multithreaded(except games and Business Winstone), even Cache/Memory portion of the Sandra 2007. Some of the apps tend to gain very little due to extra CPU but they do take advantage.

    It shows however, Core 2 Extreme has BOTH memory and cache latency advantage.

    http://www.anandtech.com/IT/showdoc.aspx?i=2772&p=4

    ScienceMark didn't agree completely and reported about 65-70 ns latency on the Opteron system and 70-76 ns (230 cycles) on the Woodcrest system. We have reason to believe that Woodcrest's latency is closer to what LMBench reports: the excellent prefetchers are hiding the true latency numbers from Sciencemark. It must also be said that the measurements for the Opteron on the Opteron are only for the local memory, not the remote memory.

    Core 2 Extreme got 36.75ns.

    http://www.anandtech.com/showdoc.aspx?i=1884&p=11

    Above we see A64 15.5% faster than AXP at Business Winstone at same clock speed. The majority of the advantage clearly lies in the integrated memory controller(there is also the advantage of doubled cache, but its a minor part of the 15.5%). The Japanese site results show 2.93GHz Core 2 Extreme performing 2% better with same hard drive than the Xeon 5160.

    I don't think what I am about to summarize is a coincidence:
    -SpecFP scores show 10% higher results for Core 2 Extreme against Xeon 5160(3046 vs. 2775)
    -Core 2 Extreme has half the latency in Sciencemark(Anandtech)
    -Core 2 Extreme only has 10% less bandwidth with HALF the memory channels compared to Xeon 5160
    -Xeon 5160 achieves 3.8GB/sec(Tomshardware) with Dual Channel DDR2-533(Core 2 Extreme achieves over 5GB/sec)
    -Japanese Site PCWatch tests show Xeon 5160 has nearly 40% higher latency in Everest
    -Business Winstone is ever slightly faster with Core 2 Extreme, with 2% clock speed disadvantage
    -The performance margins between Xeon and Opteron are somewhat narrower than Core 2 and A64/X2, Opteron has almost identical memory controller as A64 while Xeons and Core 2 has a much different one.
     
  23. DavidC1 thread starter macrumors member

    Joined:
    Jul 26, 2006
    #23
    There we go:

    Sphinx speech recognition
    http://www.techreport.com/etc/2006q2/woodcrest/index.x?pg=8

    http://www.techreport.com/reviews/2006q3/core2/index.x?pg=12

    Core 2 E is 37% faster than Xeon 5160

    also in Mental Ray Render for 3DS 7: http://www.techreport.com/etc/2006q2/woodcrest/index.x?pg=6
    http://www.techreport.com/reviews/2006q3/core2/index.x?pg=14

    Core 2 E is 34% faster


    So I may not be completely right. Just like the workstation hard drives that perform best in workstation environments, Xeon is more suited than Core 2 to a workstation environment, but not true in PC environements: http://www.anandtech.com/storage/showdoc.aspx?i=1799&p=9

    Intel did say Woodcrest/Conroe/Merom has a prefetcher optimized for their own market.

    So don't expect your everyday app/gaming to be faster on Xeon because of the faster FSB, cause it'll lose out to Core 2. But if you use workstation apps, it'll be better on the Xeon.
     
  24. Fadl macrumors member

    Joined:
    Aug 7, 2001
    #24
    Nice reviews! Thanks for that.
    I believe that the Xeon is slower in some apps cause of the FB-Dimm but you are also wrong in some parts.

    Techreport about the Woodcrest Sphinx speech recognition :"he Woodcrest system sets what I believe is a new record in this test, requiring only about 40% of a single CPU core to process this speech recognition routine in real time."

    Would they have said this when a P4 XE is faster than the Woodcrest? I think you can't compare the results from the Woodcrest/Opteron review and the Core 2 Duo reviews. If so an Opteron 285 is much slower than an A64 4200+.... ;)


    It is never good to take some benchmark results from different reviews and than compare against each other.

    But i do believe that the Xeons are slower(in some apps) than a Core 2 Duo because of the FB-Dimm which is designed to put as much ram in a system as possible. Also the X5000 chipset will(of course) be slower than a 975 because of system stability.
    But I think this is very interessting to talk about so we maybe should open a new topic for this because it has nothing to do with X3000 results.
     
  25. DavidC1 thread starter macrumors member

    Joined:
    Jul 26, 2006
    #25
    Actually I didn't want to drag it, but it sort of did, sorry about that. You can see that the latency is significantly higher on Woodcrest and also lower in bandwidth despite 2x memory channels and 25% faster FSB on the latest tests by AT on Mac Pro.
     

Share This Page