 |
 |
 |
||
|
|
||||
 |
|
|
| Introduction and Product Specifications | ||||||||
At the end of 2006, AMD announced the availability of processors built using the company's 65nm manufacturing process, which is first being implemented with the Athlon 64 X2 product line. The move to 65nm brings a number of advantages for both AMD and consumers. From AMD's standpoint, the more advanced manufacturing process allows for more processors to be produced per wafer, therefore improving manufacturing efficiency and in turn profit margins. Ultimately, the aim is to let this savings trickle down into retail channels, lowering AMD's costs and potentially the cost of the X2 lines across the board. The move to 65nm also lays the foundation for future, native-quad core processors that would be too costly to produce at 90nm. With this transition, cost is a major factor, of course, but not just from a manufacturing perspective. With the change over to 65nm, AMD has also sharpened their focus on power consumption and energy efficiency, which translates to lower operating costs as well. What AMD has not worked into the transition are any core performance enhancements. The goal was instead a more energy efficient processor that consumes less power and produces less heat, while maintaining a similar performance level as their 90nm processors. This, in turn, should translate to higher performance per watt than the 90nm design. However, since its initial release, the performance impact of changes introduced with the new 65nm CPUs has come into question.
Since its release in December, several major publications have confirmed the differences in power consumption and performance between AMD's 65nm and 90nm cores. Additionally, in testing, some have noted increased cache latencies resulting in a slight decrease in performance in some tests. In our coverage of the 65nm Athlon X2, naturally we'll compare power consumption and performance at the outlet between the 65nm and 90nm processor to quantify the differences. Then, we'll take a moment to run a series of performance tests to assess whether the higher cache access latencies have an impact on performance, or whether in the end they turn out to be much ado about nothing.
AMD's base pricing structure for the newer 65nm based Athlon 64 X2's
|
| AMD Athlon 64 X2 65nm Brisbane Core | ||||||||||||||
When the 90nm Athlon 64 X2 was released, we saw processors with a Maximum Thermal Power rating of 89 watts, with low power flavors later offering 65 and 35 watt Maximum Thermal Power ratings at lower processor speeds. With a reduced die size from 183mm2 with the "Windsor" core to 126mm2 with the new "Brisbane" core, AMD has taken the low power mantra and made it standard across the board, rather than designating a Low Power-processor and offering it as an additional cost as was done with "Windsor".
When we look at this new line up, there are several changes to consider. First, when compared to its 90nm brethren, the X2 5000+ remains unchanged with respect to L2 Cache and clockspeed, however, the rest have all shifted to a more consistent L2 complement of 512KB while adding a slight increase in clock speed.
Like the "Windsor" core, "Brisbane" maintains an approximate Transistor count of 153.8 million. The Nominal Voltage has reduced, ranging from 1.25-1.35v with "Brisbane" compared to the "Windsor" core's 1.3-1.35v. One other change is that AMD has introduced a half stepping with the CPU multiplier, allowing for even 100MHz memory frequency steps.
In both cases, the "Windsor" and "Brisbane" based Athlon 64 X2 5000+'s we are using for testing sport the same 512KB of L2 cache per core, support the same instructions and run at the same frequency. This makes for an excellent apple-to-apples test comparison. |
| HH Test Bed, Power & Thermal Characteristics (IDLE) | ||||||||||
Back when AMD introduced their Energy Efficient Athlon 64 X2 3800+ and 4600+, we did a complete comparison including processors from both Intel and AMD. In that review, it was obvious that AMD came up with an extremely efficient processor when it came to low power consumption. This review, in essence, can be considered an extension of that piece, with today's focus being an apples-to-apples comparison between the AMD Athlon 64 X2 5000+ 90nm "Windsor" and the AMD Athlon 64 X2 5000+ 65nm "Brisbane".
In our first overview of power consumption, we recorded the idle wattage of our test system with Cool'n'Quiet Enabled and with it Disabled. All measurement were recorded at 30 minute intervals using a Kill-A-Watt meter at the outlet. In all testing, a stock cooler was used along with Arctic Silver 5 thermal grease.
With Cool'n'Quiet enabled, each processor dropped down to 1000MHz from 2600MHz, which resulted in an overall system draw of 87 watts with both "Brisbane" and "Windsor" cores. With Cool'n'Quiet disabled, we start to see more of a variation between the two cores, with the "Brisbane" weighing in at 7 watts less compared to the "Winsdor" based model. Next we used ASUS Probe II to monitor the CPU temperature, once again with Cool'n'Quiet enabled and disabled, waiting 30 minutes before recording each result.
With Cool'n'Quiet enabled, both processors ran at nearly the same temperature of 18-19C. When we disabled Cool'n'Quiet, like our wattage test, we saw a slight increase in the variations, with the "Windsor" core idling 5C higher than the "Brisbane". |
| Power & Thermal Characteristics (LOAD) & Overclocking | ||||||||||
In our next test we get to the real thermal/power numbers, that being with each core pegged at 100% usage. To do this, two iterations of Prime95 were launched with the Torture test set for maximum heat load. In each test, the system ran for 30 minutes before recording the results.
With the CPU pinned at 100% utilization, we can clearly see the lower power requirements of the "Brisbane" core compared to "Windsor". Here we see a difference of 29 watts, which is actually slightly greater that the 25 watt difference between the "Windsor's" 89 watt specification and the "Brisbane's" 65 watt Maximum Thermal Rating.
When measuring the CPU temperatures during each test, we recorded a drop of 8 degrees C with "Brisbane" over "Windsor", which is a decent drop in overall operating temperature.
With respect to overclocking, we took a rather simplistic approach. In this test, all that was done was the memory was set to its lowest setting and the clock register was raised, recording the highest speed that we could post into Windows XP. The end result was an easily achieved 2.95GHz with the clock generator set for 227MHz and no voltage adjustments whatsoever. Once the system posted, we launched two iterations of the Prime 95 torture test and each test ran close to 6 minutes before throwing an exception and terminating. Surely, with a bit more tweaking of the voltage and multiplier, we could find a bit more stability. AMD Athlon 64 X2 5000+ Brisbane
|
| Synthetic Testing Analysis with SiSoft Sandra Continued - CPU | ||||||||
Now that we've got a good picture of how the "Brisbane" and "Windsor" cores compare with regards to power consumption and thermal performance, we'll shift our focus to performance.
In our first round of tests, we focused on CPU performance, doing an apples-to-apples comparison between the "Brisbane" and "Windsor" based Athlon 64 X2 5000+ processors. The first test run was SANDRA's Arithmetic Test followed by the Multi-Media test component.
When we do a side-by-side comparison of the Arithmetic test results, we see the "Brisbane" core held a small edge over the "Windsor" by 23 MIPS while Whetstone results swung the other way, with the "Windsor" core leading by 49 MFLOPS. With the Multi-Media test, we saw the 65m "Brisbane" top the "Windsor" in both Integer and Floating-Point calculations, with the Floating-Point calculations showing the widest margins at 94 it/s vs the integer test which recorded a 21 it/s difference. |
| Synthetic Testing Analysis with SiSoft Sandra Continued - Memory | ||||||||||
In our next segment, we focused on memory related performance metrics, most notably Memory Bandwidth and Memory Latency.
Overall memory performance between the two cores did show a performance hit with the "Brisbane" core. In both Integer and Floating-Point calculations, we saw the "Brisbane" trail the "Windsor" by an average of 1.5% overall. Next, we used SANDRA's Memory Latency test to assess each processor's cache and memory subsystems. With the "Brisbane" based 5000+, Random Access Memory Latency weighed in at 142ns whereas the "Windsor" based 5000+ recorded a lower 124ns latency. This equated to a 14.5% increase in latency with the "Brisbane" core, which is no small delta. To further quantify the impact of this latency on performance, we will continue our comparison between the two cores with several more synthetic and real-world benchmarks.
When we ran SANDRA's Cache and Memory Latency test, we recorded no significant variations, with the "Brisbane" coming in with slightly better results than the "Windsor" core. |
| Futuremark 3DMark06 - CPU Test and PCMark05 CPU/Memory | |||||||||
3DMark06's test is a multi-threaded "gaming related" DirectX metric that's useful for comparing relative performance between similarly equipped systems. This test consists of different 3D scenes that are generated with a software and hardware GPU renderers, which is also dependant on the host CPU's performance. In its CPU tests, the calculations normally reserved for your 3D accelerator are instead sent to the host processor.
At the conclusion of this test, we did record a performance decrease of 2.5% with the "Brisbane" compared to the 90nm "Windsor". While not as severe as the latency we recorded in SANDRA, we are seeing a slight performance slowdown with the "Brisbane" core.
"The CPU test suite is a collection of tests that are run to isolate the performance of the CPU. The CPU Test Suite also includes multithreading: two of the test scenarios are run multithreaded; the other including two simultaneous tests and the other running four tests simultaneously. The remaining six tests are run single threaded. Operations include, File Compression/Decompression, Encryption/Decryption, Image Decompression, and Audio Compression" - Courtesy FutureMark Corp.
When the PCMark05 CPU test completed, once again, we saw a slight decrease in performance with "Brisbane" that equated to less than 1% compared to the "Windsor" based 5000+.
"The Memory test suite is a collection of tests that isolate the performance of the memory subsystem. The memory subsystem consists of various devices on the PC. This includes the main memory, the CPU internal cache (known as the L1 cache) and the external cache (known as the L2 cache). As it is difficult to find applications that only stress the memory, we explicitly developed a set of tests geared for this purpose. The tests are written in C++ and assembly. They include: Reading data blocks from memory, Writing data blocks to memory performing copy operations on data blocks, random access to data items and latency testing." - Courtesy FutureMark Corp.
With memory specific performance, PCMark05's Memory module posted a 9% decrease in memory performance with the 65nm "Brisbane" vs the 90nm "Windsor" core. This is a fairly major drop in our opinion, however, we'd like to see how this translates in real world application testing before drawing any conclusion. |
| World Bench 5.0: Office XP SP2, Photoshop 7 & MT Modules | ||||
To see how the synthetic performances we've seen thus far translate into real world performance, we ran several common tests using PC World Magazine's WorldBench 5.0. This Business and Professional application benchmark suite consists of a number of performance modules that each utilize one, or a group of, popular applications to gauge performance. In this segment, we ran Office XP SP2, Photoshop 7 & MT Modules.
Office XP SP2 testing actually showed the "Brisbane" core completing the test six seconds faster than the "Windsor" based CPU, which equals 1.2% overall.
Opposite of the Office XP SP2 test, the Photoshop module reported our Athlon 64 X2 5000+ "Windsor" as the quicker CPU, topping the "Brisbane" by 1%.
WorldBench 5's Multithreaded test showed the widest margins seen in real world testing thus far, with the 90nm "Windsor" completing the test 11 seconds faster than the "Brisbane". This difference measure up to 2.5% overall. |
| LAME MT MP3 Encoding & Quake 4 Low Res. | ||||||||||
In our custom LAME MT MP3 encoding test, we converted a large WAV file to the MP3 format. In this test, we created our own 223MB WAV file (a never-ending Grateful Dead jam) and converted it to the MP3 format using the multi-thread capable LAME MT application in single and multi-thread modes.
Processing times are recorded below. Once again, shorter times equate to better performance.
With our LAME MT testing, we saw both processors compete on the same level, with a mere one second difference noted in Multi-Thread testing, which does equal a 2% difference and is consistent with several variations previously reported.
In our final test, we used our custom Quake 4 demo to assess subsystem performance by setting the resolution to 640x480 and the image quality to low. The end result was the "Windsor" core offering 3 FPS more than the "Brisbane", a lead of 3.48%. |
| Performance Summary and Final Thoughts | ||||
Performance Summary: In regard to the thermal and power characteristics of the new 65nm "Brisbane" core, AMD seems to have delivered on their claims of lower power consumption, which in turn translates into a higher performance per watt ratio than previous 90nm AMD processors. Strictly from a performance standpoint, however, the higher cache access latencies of the Brisbane core did translate into somewhat lower performance overall. With Synthetic testing, we did note lower performance in PCMark05's Memory test and SANDRA's Memory Latency Test. However, when we shifted our focus to performance testing with real-world applications, the actual differences reported were about 1-2.5% across the board.
From what we've seen in our testing, AMD delivers on their promise when it comes to a lower power CPU. In regards to both wattage draw and thermals, the new 65nm CPU was better overall than the same speed processor based on the older 90nm "Windsor" core. Lower power also comes with slightly lower performance in the case of AMD's 65nm Brisbane core, however. In synthetic testing we saw the widest performance deltas, some of which hovered around 14.5% compared to the older CPU. Yet, when we shifted to real-world testing, the margins were a more marginal 1-2.5%. In the end though, slower obviously isn't better. However, in a side-by-side comparison, we think users would be hard pressed to actually "feel" the difference between the two processors. In fact, during our evaluation of the 65nm AMD Athlon X2 5000+, we ran the CPU in a normal work environment for several days and then switched it out with its 90nm counterpart and there was no obvious difference is day to day performance. In addition to this, the 65nm Brisbane core-based processor overclocked much better than the Windsor-core. And the higher overclock would easily offset the higher latencies. If you're in the market for a mid-range Athlon 64 X2 processor and plan to do some overclocking, look to the newer 65nm processors. Otherwise, if you're planning to run an AM2 system in an all-stock configuration one of AMD's 90nm processors offer slightly better performance.
|
;)
;)
;)
;)
;)
;)
;)
;)
;)
;)
;)
;)
;)
;)
;)
;)
;)
