3GHZ Woodcrest against 2.6GHZ Opteron performance duel
Under the watchful eyes of Intel employees, TechReport did some benchmarks of the Woodcrest. The results show that Woodcrest 3GHZ is only slightly faster than 2.6GHZ Opteron 285 in most tests. In multi-threaded applications, Woodcrest is often only 5% faster, despite having a 15% clockspeed advantage. Please note that the tests weren't even server benchmarks where bandwidth is crucial, they were Windows desktop applications.
My previous analysis reached similar conclusions. As you can see from TechReport's POV Ray benchmarks, the Woodcrest/Conroe has quite some advantage in single threaded loads, where one core can use the whole 4MB shared cache. However, once you start 2 threads for 2 cores - in other words, each core can only use 2MB of cache, Woodcrest's performance lead reduces to 6%. Since the Woodcrest engineering sample has a 15% clockspeed advantage, this indicates that the CORE2 is 10% slower than K8 at the same clock. This is in agreement with my previous analysis, that CONROE/Woodcrest/Clovertown only has an advantage for single threaded loads that have a working set comparable to 4MB.
TechReport also reports that the Woodcrest server platform isn't entirely stable yet. Woodcrest suffers BSOD (Blue Screen Of Death) under certain conditions. Woodcrest is set to launch on June 19. Intel better hurry.
Realtech.com also performed a set of benchmarks on the 3GHZ Woodcrest, the results on Cinebench 9.5 and POV Ray were in agreement with TechReport's. Realtech.com obtained a SPECjbb2005 result of 56,835 BOP/s. In comparison, a 2.4GHZ Opteron obtained a score of 62074 BOP/s (under Linux).
In March, answering the question on whether DELL would go AMD, DELL CEO Kevin Rollins said:
"There are multiple decision points. Is Intel going to meet the technology needs-–server performance and thermals–-where AMD does have a lead? That will answer the question. If they don't, that will also answer the question."
Intel is kinda weird. Today, it launched the Dempsey server CPU, just 3 weeks ahead of the Woodcrest. The 3GHZ Xeon 5050 dual core is only $177. Dempsey is dirt cheap, actually, cheaper than dirt. I admit that even I am tempted by this part. But once you look at the total system cost, it is actually quite expensive. The motherboard alone costs about $800. Adding the electricity bills, the equation changes completely.
posted by Sharikou, Ph. D. @ 5/23/2006 10:20:00 AM
28 Comments:
Wow...I just read the review, and 1 thing came out:
Woodcrest was about 5-15% faster, given 400MHz Clock speed advantage, DOUBLE the Memory Bandwidth (+9GB/s if that, vs. 17GB/s) and this is a next-gen chip? Come on people!
Repost the BSOD with verification from the highest peak in the land. The stability issues and lack of quailification of the "platform" are the major issues in the server area. AMd qualification process took 1 yr. Intc hasnt qualification hasn't been even one month as this cow cake of a processor is being rushed to market. Start printing erratta list for the termite infested woody. Lack of stability of woody is THE major issue!!
"Woodcrest was about 5-15% faster, given 400MHz Clock speed advantage, DOUBLE the Memory Bandwidth (+9GB/s if that, vs. 17GB/s) and this is a next-gen chip? Come on people"
Based on recent AM2 DDR2 review, K8 doesn't benefit from memory bandwidth. I am wondering when Opteron can break 3Ghz.
"the Dual-Core Intel Xeon processor 5100 series, previously codenamed "Woodcrest" -- delivers up to 125 percent performance(2) improvement over previous generation dual-core Intel Xeon processors and up to 60 percent performance(3) improvement over competing x86 based architectures(4), whilst also delivering performance per watt leadership.(5)"
What benchmarks are those? SuperPI and Woodcrest Idle vs. Opteron Full Load?
"Based on recent AM2 DDR2 review, K8 doesn't benefit from memory bandwidth. I am wondering when Opteron can break 3Ghz"
Again, what were you looking at? Intel benchmarks?
Intel basically admitted that their Xeons today are 60% slower than Opteron. Their promise is Woodcrest 3GHZ will be faster. But look at HP's system availability date, the Woodcrest servers won't be available until November 22, 2006.
"But look at HP's system availability date, the Woodcrest servers won't be available until November 22, 2006"
Manufacturing takes time, isn't that obvious???
"Intel is kinda weird. Today, it launched the Dempsey server CPU, just 3 weeks ahead of the Woodcrest."
That's what happens when one product gets delayed repeatedly, and its successor comes in ahead of schedule. Intel aren't the only people who do this though; ATI did it with the Radeon X1800 and X1900.
Woodcrest benchmark are not so bad after all, but AM2 introduction is a real disappointment. They should at least try to put Z-RAM and L3 cache to improve performance...
Isn't it wierld that Woodcrest, clock-for-clock, isn't better than Opteron, but Conroe (again clock-for-clock) seems better than Athlon64 (see Hexus latest review)?
Isn't it wierld that Woodcrest, clock-for-clock, isn't better than Opteron, but Conroe (again clock-for-clock) seems better than Athlon64.
The difference is 32 bit and 64 bit. Intel runs slower in 64 bit. AMD runs 20-40% faster in 64 bit. If you look at the 64 bit applications, 3GHZ Woodcrest is often only 5% faster than 2.6GHZ Opteron, which means Woodcrest is actually 10% slower than Opteron at the same clock. AMD's 64 bit performance is native and can more than compensate for the much smaller cache.
Wow Intel will make real fast gaming PC's duh we want servers for virtualization and scalability what the hell is Intel playing at, AMD is killing them in the business were money is made.
Yonah was a good damage-control move by Intel, getting dual-core chips to the mobile market earlier and offering tangible performance improvements in mobile 32-bit computing. I wish that AMD had made a larger move with their first AM2 chips... If they had incorporated a couple more features of the K8L architecture into their new AM2 chips, I'm sure the early AM2 chips would be completely uncontestable.
This is Chicagrafo here.
Something that may be of interest to you and your audience:
Focus on the Blue Screen Of Death issue. I predicted that the complexity of Core2 would lead to buggy processors and delays, probably, Intel won't have a choice but to implore Microsoft to come up with patches that go around the bugs in the processors. Expect the errata pages to be very long.
I am relieved for the confirmation of the prediction about bugs. I still think that it will be extremely difficult for Intel to iron out all the bugs; their track record is preposterous
I'm pretty impressed by Woodcrest, actually. It beats AMD in nearly all the benchmarks by a decent margin.
"I am relieved for the confirmation of the prediction about bugs. I still think that it will be extremely difficult for Intel to iron out all the bugs; their track record is preposterous"
I think this will screw Intel's reputation big time, especially with server customers where credibility means a lot.
"I'm pretty impressed by Woodcrest, actually. It beats AMD in nearly all the benchmarks by a decent margin."
Did you consider the variables in the test? Did you consider what will the results be if it was done in 64 bit?
"Did you consider what will the results be if it was done in 64 bit?"
Many of the benchmarks were done in 64-bit, and Intel's Woodcrest still won by the same margins as the 32-bit apps.
Again, I note that Intel will not show results of TPC and SPEC (except SPECint) benchmarks. Instead, they will only allow benchmarks which are not particularly relevant to server performance. My guess is that not too many servers are sold to run single threaded ray tracing or encoding apps.
Based on my analysis of Intel's own claims back at Spring IDF, I did NOT expect Woodcrest to outperform Opteron in anything except pure integer apps in a 2P configuration. That was also assuming that Opterons topped out at 2.8 GHz, which is already OBE.
http://www.xbitlabs.com/news/cpu/display/20060411063742.html
Everything we see and everything we don't see points to one conclusion: Opteron will still be better in performance/watt at 2P and up.
Again, I note that Intel will not show results of TPC and SPEC (except SPECint) benchmarks. Instead, they will only allow benchmarks which are not particularly relevant to server performance. My guess is that not too many servers are sold to run single threaded ray tracing or encoding apps.
From here you get TPC-C, SPEC rates, SPECjbb, SPECWeb.
http://www.intel.com/performance/server/xeon/database.htm
http://www50.sap.com/benchmarkdata/sd2tier.asp
has SAP-SD
how about using a sata II drive for the intel machine and a sata I drive for the opeteron box? how about them stating the dirver of 3/14/6 when tyan's website shows the current drive r to be of 3/27/06?
http://www.tyan.com/support/html/b_s2895.html
Concerning the POVRay test at http://www.techreport.com/etc/2006q2/woodcrest/index.x?pg=6 did you read between the lines? According to the Tech Report reviewers, "Once more, Woodcrest comes out on top of the Opteron. The margin of victory here appears to narrow as we add threads, but I should be clear: that is no fault of Woodcrest. The newest version of SMPOV takes more time than usual spinning off new render threads, and I'm certain that's why neither system scaled as well as it could have.". Don't jump to your own conclusions so quickly in the future. Using a single benchmark for your theories (as you had done earlier) does not always have a good basis.
another thing besides the use of a lower speed hard drive in the benchmark setup is the memory. they indicate 8 dimms for 4GB. but wait, there was only one opteron 285, so that means all available 8 dimms were occupied. but all 8 dimms don't get used unless you put another cpu in the second socket! so really they had 2GB for the opeteron and 4GB for the intel machine. also what timings? kington has several different DDR400 ECC memories for the tyan K8WE board.
Not sure if this has already been posted, but never mind:
Link
PS: The benchmarks also include the Opteron 254 in single threaded tests where the Woodcrest also uses 1 core only.
>SMPOV takes more time.
The actual version should be even faster then that.
Anyway, when doing rendering-tests with SMPOV its really important, to choose pictures having a long render-time.
As its said in my benchmark-descriptions, in that case you have only a fraction of time for the SMPOV-Overhead.
Choosing pictures, with a short render time and large tile numbers may not even work well with the actual version.
Also the new version (while beeing a preview version) has one major change, tiles are now choosen by a random algorhytm.
the reason for this IS that the render-jobs will start faster.
That means the rendering-order in two runs is with the actual version never equal.
Later versions may have a benchmark-mode (its not yet implemented).
If you choose pov-files with long rendering-times e.g. benchmark.pov 1600x1200 etc. the results should stand a later verfication and should be reproducable. Unless in a Network with diffrent CPU's.
One more remark: For benchmarks, do nor use the Preview-Feature.
It uses actually up to 30% Performance if not set on low prio, as its only for demonstarion purposes.
Sharikou, why are you trying to spread disinformation wrt the SECjbb2005 results from realworldtech.com?
Their 56,835 result for 2-sockets of 3.0GHz Woodcrest was obtained on a 32-bit JVM (unspecified vendor, but probably a Sun JVM).
The SPEC.org result of 62,074 for 2 sockets of 2.4GHz Opteron 880s was obtained on 64-bit JRockit5.0 engine from BEA.
It's completely ludicrous to attempt any sort of comparison between results obtained at different bit-widths and on different VMs. And you're smart enough to know it.
Their 56,835 result for 2-sockets of 3.0GHz Woodcrest was obtained on a 32-bit JVM (unspecified vendor, but probably a Sun JVM).
The SPEC.org result of 62,074 for 2 sockets of 2.4GHz Opteron 880s was obtained on 64-bit JRockit5.0 engine from BEA.
The RealWorldTech benchmark result was obtained under BEA JVM too (you can see from the cmd line args). Based on a statement by a BEA developer, 64bit JVMs are slower than32 bit ones by 10-20%. You can read more about the specjbb2005 results there.
Quote:
"Independent results show Intel's Woodcrest 10% slower than Opteron at the same clock under 64 bit mode."
Intel:
Xeon DP 5110: 1.60 GHz, FSB1066, 4 MB L2 cache, $209 at launch
Xeon DP 5120: 1.86 GHz, FSB1066, 4 MB L2 cache, $256 at launch
Xeon DP 5130: 2.00 GHz, FSB1333, 4 MB L2 cache, $316 at launch
Xeon DP 5140: 2.33 GHz, FSB1333, 4 MB L2 cache, $455 at launch
Xeon DP 5150: 2.66 GHz, FSB1333, 4 MB L2 cache, $690 at launch
Xeon DP 5160: 3.00 GHz, FSB1333, 4 MB L2 cache, $851 at launch
source: http://www.tgdaily.com/2006/05/23/intel_woodcrest_benchmarks/
AMD:
Dual-Core Model 265 (1.8 Ghz) $316
Dual-Core Model 270 (2.0 Ghz) $455
Dual-Core Model 275 (2.2 Ghz) $690
Dual-Core Model 280 (2.4 Ghz) $851
Dual-Core Model 285 (2.6 Ghz) $1,051
source: http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_609,00.html?redir=CPT301
In other words:
$851: Opteron 2.4 Ghz - Xeon 3.00 GHz
(+25% clock speed)
$690: Opteron 2.2 Ghz - Xeon 2.66 GHz
(+21% clock speed)
$455: Opteron 2.0 Ghz - Xeon 2.33 Ghz
(+16.5% clock speed)
$316: Opteron 1.8 Ghz - Xeon 2.0 Ghz
(+11% clock speed)
No further comments...
Remember The P3 1.13 debacle. paper launch and then recaled
Dell is shipping Woodcrest in July 2005. Just check their webpage www.dell.com
Post a Comment
<< Home