AMD explains Shanghaied benchmarketing
AMD CONFIRMED the benchmarks are Shangahi's but complained about our analysis of the numbers.
Here's it's statement:
The goal of these submissions was to enable comparisons of comparably configured servers , aka apples to apples in what you see in a mainstream, typical configuration.
· Each server uses 8 DIMMs of 2GB each. This is a typical server configuration which uses 2GB per core. There are plenty of users out there who put 8 DIMMs in a server probably would be interested in the score. In fact, it might be worth your time to see how many people use 8 DIMM servers vs. the San Clemente servers that the top Intel scores are based on. Or at least a good sidebar discussion to have.
The Intel score uses Low-Power FB-DIMM ("Memory Details: DDR2-667 CL5 LP
FB-DIMM;") If we were trying to create a "worst case scenario" as you call out
below, we would have used standard power FB-DIMM memory in the Intel server.
· Each server uses the same Ablecom PWS-702A-1R power supply
· Each server uses the same 500GB WD5000ABPS SATA II disk drive
· Each server uses the same Intel Pro/1000 PT Dual Port Server Adapter NIC
· Each server uses "Dynamic Power Profile Enabled"
Again, if we were trying to show worst case scenario we would have turned
off Power Management on the Intel server
· Each server uses the same JVM version
· Each server uses the same JVM command-line options
AMD-supplied L'INQs
731
score using 2.5GHz AMD Opteron 2380 in Supermicro 2021M-UR+ with Ablecom
PWS-702A-1R power supply.
561
score using 2.5GHz Intel Xeon L5420 in Supermicro 6025W-NTR+ with Ablecom
PWS-702A-1R power supply.
Share this:
ShareBusiness Intelligence Software
Customer Relationship Management (CRM)
Enterprise Accounting Software
Enterprise Resource Planning (ERP)
Enterprise Systems Management
http://www.spec.org/power_ssj2008/results/res2008q4/power_ssj2008-20081017-00089.html

What's interesting in AMD's 2.5GHz L5420 submission is that under the "System Under Test Notes" they state this "Hardware Prefetch disabled in BIOS (default enabled)". If I'm not mistaken, hardware prefetch should increase performance since it's an important mechanism to ensure Harpertown can take advantage of it's big caches. And even if that weren't the case, disabling an option that is enabled by default goes against AMD's claims that they are trying to set up a default build as what is available in the market.

The other discrepancy is that AMD set the Intel L5420 systems with a JVM heap size of 3600MB while their own Shanghai has a 3700MB. I'm not sure what that does, but it is different between the systems.
Well, people would also be "interested" in the score of 2x8 instead of 4x4 for the AMD SYSTEMS also, because that would make an "apples to apples" comparison.

Freakin' CHEATERS.
I regards to my earlier comments on AMD disabling hardware prefetch, after looking at other SPEC submissions, it seems that it's a common occurrence in Intel submissions, so I guess AMD was justified in doing so. Intel submission also usually have Adjacent Cache Line Prefetch disabled in the BIOS as well, which AMD didn't do. I'd be curious to know what these BIOS options do and how they affect performance in SPEC.
I noticed that both the high scores - 1135 for Intel and 860 for AMD use the -XlargePages JVM flag which is missing from the setup used to obtain the 561 score. Could this be the reason for the low score .... If so, shame on AMD ....
Coming from AMD themselves the score will be scrutinized against other known scores. Many holes to pick with this one like the large disrepency.

This is going down in infamy along with the interpolated Barcelona benchmarks, and the canned reviews on Barcelona launch.
Hi-

Just wanted to let you know, the AMD guy does have a point regarding the DIMM layout. When configuring machines with a large RAM configuration, the price difference between filling all of the RAM slots with smaller dimms and using only a couple of slots, but with large DIMMs is very significant.

From your article, it appeared there is a 2x performance difference when using a full set of dimms and only a couple. It might be interesting to do an article where a side by side comparison is made of an amd and intel box with each having only a couple of dimms and a full set of dimms, yet with the same total RAM count.

Quite probably comparing the AMD and intel will be like apples and oranges due to the architecture diffences. In the end though, we the buyers don't care so much for the how and whys, as we care for price/performance.

example1: need to build a vmware server with 24gb ram.
example2: need to build an OLTP database server for maximum transaction count processing
example 3: need to build a data warehouse server for analytical research.

Which combo could be more effective?

Cheers

Elmars
The difference is very simple to explain if you compare the power at idle load. (with the same power supply)
AMD= 138W
intel= 179W (+30%)

At Max. Load:
AMD=269W
intel=279W (+4%)

This big difference at the start is costing big in total.
AMD has always been better in turning the power down when not needed.
And knowing that just the low load wattage is a big factor in estimating your power bill AMD is the winner here. (even if we assume a better speed score for intel than shown)



-Xlargepages is implicitly enabled by the use of -XXaggressive in the submissions. No dirty trick there.

You can find all details here:
http://www.spec.org/power_ssj2008/results/res2008q4/

The key observation is that the difference between the AMD Xeon score (561) and the Supermicro Xeon score (990) is small in terms of actual performance (peak 268k vs 279k -> 4% difference) which implies that AMD didn't play any real tricks to make Xeon underperform. The difference is instead in the power consumption: 179-279W for AMD's submission vs 97-174W for the Supermicro submission.

Whether the comparison is "fair" or not depends on whether you agree with AMD in that this is a reflection of real-world configurations.

Full disclosure: I work in the Oracle JRockit team. We work with Intel and AMD but have not been involved in these benchmark submissions.

Cheers,

Henrik
FBDIMMs use more power than DDR2. Imagine 8 sticks of FBDIMMs... This ain't apples to apples!

AMD are big cheaters.
It's surprisingly simple: DO NOT EVER BELIEVE ADVERTISEMENT.
And never believe tests done by the parties themselves.