backtop


Print 77 comment(s) - last by ditkaamdVSinte.. on Jun 2 at 6:05 PM

DailyTech takes a look at Intel and AMD's claims when it comes to server processor performance, pricing.

With the recent release of Intel's Nehalem-based Xeon processors last month and AMD's Shanghai-based Opteron processor in November 2008, a war of words has precipitated between Intel and AMD.  Intel claimed that it maintained the greatest performance-per-dollar and that its slowest Nehalem was faster than the fastest Opteron, while AMD refutes this and claimed that it provides a better value product.  AMD also claims to have the most energy efficient chip by advertising the world's most efficient and first "40-watt" quad-core CPU while Intel claims that its CPUs are the most energy efficient. 

The result of this public bickering is that it is difficult for chip customers to figure out who to believe.  To settle this dispute, we will analyze each of these claims and provide hard data to validate or debunk each of these claims.

Claim #1:
The slowest Intel Nehalem is faster than the fastest AMD Shanghai Opteron processor.

Reality: No, yes, and not anymore.
If we include the one dual-core Nehalem variant, the answer is no because a lower clock speed dual-core Intel Nehalem isn't going to beat the highest clocked quad-core AMD Shanghai.  If Intel was only talking about all of the other quad-core Nehalem models then yes, even the slowest 2 GHz Nehalem XEON model E5504 will soundly beat the fastest 2.7 GHz AMD Shanghai Opteron model 2384.  But this has changed recently as AMD has just recently released a 3.1 GHz model 2393SE Opteron which would likely be faster than the slowest Nehalem quad-core processor but still lag far behind most Nehalem model numbers.

Table 1a and 1b below shows the latest and best performance scores published at SPEC.org as of 4/27/2009.  SPECint is a general purpose benchmark covering a lot of mainstream applications while SPECfp represents a cross section of scientific and engineering applications.  SPECweb reflects web server performance.  For comparison sakes, we'll also include the fastest Nehalem Xeon 2.93 GHz X5570 which on average is nearly twice as fast as the Opteron 2384.  Since official public benchmarks aren't available at the time of this writing, table 1 below will not include model 2393SE performance numbers.  However, the performance gain of the Opteron 2393SE will likely be no more than 15% faster than an Opteron 2384 which may be enough to put it past the Intel E5504.

Table 1a: SPEC CPU 2006 base performance results as of 4/27/2009

Processor

Chip price

GHz

Cores

SPECint

SPECfp

Xeon E5504

$235

2.0

2x4

127

111

Opteron 2384

$957

2.7

2x4

113

105

Xeon X5570

$1470

2.93

2x4

241

194

The SPECfp data is especially noteworthy because this was traditionally an AMD stronghold before the arrival of Intel's Nehalem processor with QuickPath memory architecture.  SPECfp along with everything else is now clearly dominated by Intel Nehalem.

Table 1b: SPEC WEB 2005 performance results as of 4/27/2009

Processor

Chip price

GHz

Cores

SPECweb

Opteron 2384

$957

2.7

2x4

39,530

Opteron 8384

$1865

2.7

4x4

48,007

Xeon X5570

$1470

2.93

2x4

71,045

SPECweb was another strong suit for AMD before the arrival of Intel Nehalem.  Now a dual-socket Nehalem based Xeon server can even outpace a four-socket Shanghai based Opteron server.

Claim #2:
AMD launched the world's first quad-core processor within a 40-watt thermal envelope and has the most energy efficient processors.

Reality: False
AMD did not release a quad-core processor within the 40-watt thermal envelope and they do not have the most energy efficient chip.  While AMD never officially claimed that their newest Opteron 2377HE chip runs inside a 40-watt thermal envelope, this is the impression they have successfully marketed to the public.  AMD now refuses to disclose the actual Thermal Design Power (TDP) of their latest CPUs to the media and has tried to state that AMD's Average CPU Power (ACP) metric is comparable to Intel's TDP metric.  As a result, every news story, every product label, and every server vendor only show ACP numbers and not the actual thermal ceiling of the processors while Intel continues to use the TDP metric which more accurately reflects peak power consumption.

AMD also claims that Intel's processors have a higher thermal ceiling which makes them less suitable for dense data center deployments, but even this claim is false for most Nehalem models.  The most comprehensive and standardized server power efficiency metric to date is SPECpower_ssj2008 and it shows a wide range of power consumption metrics across a variety of load levels.  Table 2 below will show a comparison of the most efficient AMD Shanghai based servers against the most efficient Intel Nehalem based servers.  SPECpower_ssj2008 is so interesting because it not only tells us the actual measured idle and peak power consumption under a server side java load, but it also tells us the server's server side java performance.  The "score" is determined by an average measurement of performance per watt across all the workloads.  SPECpower scores for the 2.0 GHz E5504 were not available, so the closest model number with the most similar clock speed and power consumption characteristics were used in table 2.

Table 2: SPECpower_ssj2008 results as of 4/27/2009

Processor

Chip price

ACP

TDP

GHz

Cores

Score

SSJ_OPS

Measured system power

Peak

Idle

Xeon L5520

$560

 

60 W

2.26

2x4

1813

439,831

170 W

66.3 W

Xeon X5570

$1470

 

95 W

2.93

2x4

1860

620,270

244 W

81.4 W

Opteron 2376HE

$500

55 W

79 W

2.3

2x4

1044

346,326

210 W

119 W

Opteron 2384

$957

75 W

95 W

2.7

2x4

860

338,577

264 W

121 W

As we can see from the official SPECpower_ssj2008 results above, AMD's "High Efficiency" (HE) products actually consume substantially more power than Intel's low power products despite the fact that AMD advertises the lowest power consumption.  The idle power consumption gap is even larger which is significant since most servers spend most of their time low utilization states.  It's clear that the million plus transistors Intel devoted to power management on the Nehalem chip is paying off.

While these measured peak power numbers don't reflect the maximum power that each system can draw under any workload, it does represent the peak power consumption under this particular type of workload which is generally not as intensive as a High Performance Computing (HPC) workload used in scientific and engineering applications.  What these numbers do confirm is that AMD's claim that their ACP metric is most similar to Intel's TDP metric is clearly false and that AMD's TDP ratings for these processors are far more comparable to Intel's TDP.

It's also noteworthy that this benchmark also shows lower clocked Intel processors performing more server side java operations per second (SSJ_OPS) than the higher clocked AMD processors which further reinforces the data from Table 1.

Claim #3: An equivalent Intel Nehalem server costs almost twice as much as an equivalent AMD Opteron Server.

Reality: This claim is false because there are no equivalent (in performance or power consumption) AMD Opterons compared to Intel Nehalem servers.  While AMD based servers are the cheapest, a low end Nehalem server can deliver more performance at a lower price.  Table 3 below shows server prices from HP with the base configuration plus second processor.

Table 3: Server price comparison

Server name

CPU

Cores

GHz

Memory

Server Price

HP DL380 G6

Intel E5520

2x4

2.26

6 GB

$3,448

HP DL380 G6

Intel X5550

2x4

2.66

6 GB

$6,039

HP DL385 G5p

AMD 2382

2x4

2.6

8 GB

$2,619

HP DL385 G5p

AMD 2384

2x4

2.7

4 GB

$4,039

It appears that AMD is already heavily adjusting their processor prices to the major server makers rather than the unrealistic list prices that were originally set before the arrival of Intel Nehalem.  In fact if we only went by the official processor list prices, we would expect the AMD servers to cost much more than they do in table 3.  It is normal to see a lag in price adjustment immediately after a new chip lunch when customers haven't all caught on to the new products yet and when inventory is still working its way into the channel.  However, we should expect more AMD price cuts to come soon because there's no reason that a server based on an Opteron 2384 should fetch more money than a server based on Intel Nehalem server given the benchmark results.

Implications on the high end multi processor server market

The implications of Intel Nehalem on the server market for the near term are profound.  It not only threatens AMD in the two-socket market, but it threatens AMD's four-socket market and even begins to encroach on Intel's four-socket Dunnington servers.  When you can buy a two-socket Intel Nehalem server at a fraction of the price and exceed the performance of four-socket AMD servers and come close to Intel Dunnington, it takes the thunder out of four-socket servers.  Table 4a and 4b shows two-socket Intel Nehalem servers beating Opteron four-socket servers in SAP and Virtualization performance.

Table 4a: SAP performance *

Processor

Chip price

GHz

Cores

SAP

Opteron 8384

$1865

2.7

4x4

440,000

Xeon X5570

$1470

2.93

2x4

510,670

 

Table 4b: VMware virtualization performance *

Processor

Chip price

GHz

Cores

VMmark

Opteron 8384

$1865

2.7

4x4

20.35 @ 14 Tiles

Xeon W5580

$1707

3.2

2x4

23.96 @ 16 Tiles

These results are stunning because at no time in the last 5 years has the performance gap between Intel and AMD been so large.  This situation won't change until AMD launches 6-core Istanbul Opteron processors which are expected by June 2009 and Intel launches its 8-core Nehalem-EX processor for multi-socket severs expected by the end of Q4 2009.  AMD 6-core Istanbul will narrow their performance deficit not enough to close it against Intel current Nehalem-EP 4-core processors.  However, Intel's Nehalem-EX 8-core processor is expected to widen Intel's performance lead by a significant margin by the end of this year.  But because Nehalem-EX won't be seen in most mainstream two-socket servers, Istanbul will allow AMD to raise their average selling price to a healthier level.

Decision time

It appears that the server market between Intel and AMD is now mirroring the desktop chip wars.  IT customers have a choice of buying low-cost value servers based on AMD Shanghai processors or they can buy premium priced Intel-based servers for better performance and energy efficiency.  However, the value hunter should wait a few more weeks when Shanghai servers settle down to post-Nehalem prices.

* All results and prices as of 4/27/2009



Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: Not Suprised...
By psychobriggsy on 5/5/2009 7:49:22 AM , Rating: 2
Some of the writing in this article is straight from Intel's marketing handbook. The rubbish about TDP for example - AMD used to use TDP (Voltage x Amperes) whilst Intel use an achievable (thermal bug) value which is around 75% of the aforementioned value. AMD switched to their new system to match Intel's definition. I think that Intel's definition is reasonable to use, but don't go and start claiming that AMD's redefinition to something similar is somehow dishonest!

Also the scaling on the SPECWeb results for the quad CPU Opteron is highly suspect given that AMD's platform scales very well (and indeed there is no quad-socket Nehalem offering currently). Of course Nehalem performs excellently, but something is really fishy with the scaling for AMD's results, and I suspect that there are platform differences. Again this could simply be because AMD need to move to HT3 on their server platform.


“We do believe we have a moral responsibility to keep porn off the iPhone.” -- Steve Jobs











botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki