Intel VS. AMD : Performance
Performance
Now that we have the setup for all the benchmarks shown below,
let's get on with business.
The first benchmark is POV-Ray 3.50c which is a very floating-point intensive application that creates 3D images by processing a scene description from a configuration file. Our benchmark is described on POV-Ray's site on their benchmark page.
In our first look at 64-bit Linux, we used POV-Ray and noticed that not only was the AMD64 architecture much faster at this application, but that due to 64-bit verses 32-bit precision, it produced a truer image. Above we see that that Athlon 64 3800+ runs away with this benchmark under both 32-bit and 64-bit Linux. In 32-bit Linux, the Athlon 64 won over the next closest Pentium 4 by 14% and in 64-bit Linux, it won by 33%. Looking only at Intel's platforms, we see that the extra cache on the Extreme Edition chips was of bigger benefit than clock speed. We also see that the lower latency DDR memory of the 875P chipset won out by just a hair.
Next we take a look at audio encoding by using our favorite open-standard audio compression, Ogg Vorbis. Much like the common MP3 format, Ogg Vorbis is a compressed audio format which aims at compressing audio to save on disk space while keeping the sound quality as high as possible. We used the command-line encoder oggenc to encode our Mozart: Allegro Vivace track (11 minutes 22 seconds). We used the default settings with no arguments and also specified a bit-rate of 192 for testing.
Above you can see that the Athlon 64 wins not only in 64-bit Linux but also in 32-bit Linux. In 32-bit Linux, the margin was approximately 10% and in 64-bit Linux it was about 33%. This is another huge margin for AMD when using a 64-bit Linux OS distribution. All three Intel chips/platforms were almost identical, with only 10ths of a second separating them
In the next benchmark we evaluate which processor can compile a rather large application under Linux. This benchmark used to be performed with the Linux kernel but, due to the code differences between compiling for AMD64 or x86, we had to find a new application. We chose Mozilla Firefox due to it's popularity and code size. As you'll see, this application take almost 30 minutes to compile on either platform. For compiling, we used Gentoo's emerge utility which also takes into account the time it also takes to unpack the source to disk. For our make flags, we ran this with “-j#” where “#” is number of processors (real of virtual) plus one. This means that on the Intel platforms, this was performed with “-j3” and on the Athlon 64 it was performed with “-j2”. The “-j#” option tells make to compile with # of instances, increases the system's ability to use all available processing power.
As expected, the Intel Extreme Edition processors which has an additional 2MB of L3 cache slightly beat out the higher clocked Prescott. Also, we see that the 875P ran just a hair faster than the 925X due to it's lower memory latency. Athlon 64 in 64-bit mode was again faster by quite a bit verses the fastest Intel part. Here though, in 32-bit mode, the Athlon 64 fell behind and all Intel platforms won out.
From here, we move on to graphic-centric applications. SPECViewperf is a professional graphics benchmark application that aims at reproducing how actual applications will perform. SPECViewperf 7.1.1 was used here but we will be moving to the newly released 8.0 version in the future.
In this benchmark we see that Intel is a consistent leader with their highest clock speed processor on all but one test. Coming in second is the Athlon 64 under 64-bit Linux, followed by the Athlon 64 under 32-bit Linux, and finally the two Pentium 4 Extreme Editions.
Next, we have the latest first-person shooter to be released for Linux, Unreal Tournament 2004. In this benchmark we have the game engine play through a match with several computer-controlled players for a determined amount of time and report on the average frames per second over that period. You can find further instructions on how this test was performed and the files needed to duplicate this test yourself in this file that we've put together: lhoutbench-01.tar.gz
Above it's clear that the Athlon 64 is a strong gaming processor, as it edges out the Intel processors no matter whether you run it in 32-bit or 64-bit mode. The 64-bit Linux version is a bit faster but not by much. The Intel Extreme Edition chips lead over the 200MHz faster 3.6GHz Prescott. It seems that UT2004 either prefers more cache or just the Northwood core (or both).
Next we throw in the very old but seasoned Quake 3 benchmark. Here we used the built-in timedemo benchmark to see which processor will run Quake 3, and possibly other games based on the same engine, faster. We used the last version of the demo released, 1.11, which can be found in various places on-line or as Gentoo ebuild quake3-demo. Note that this is the only benchmark in which we run a 32-bit binary on our 64-bit Linux install. There isn't currently a 64-bit version of Quake 3.
As we've seen time and time again, this game just favors the Pentium 4 architecture. We suspect that it was originally compiled with those optimizations and that's why it falls behind on the Athlon 64 processor. This benchmark will only really apply to those still playing Quake 3 engine games and there are plenty of you out there. Although the Intel chips are faster here, it's hard to care when at the highest resolution, you're still above 350 FPS!
Our final set of benchmarks evaluate a couple of on-board components. The next two graphs show the average bandwidth we were able to obtain from the on-board Ethernet chips and the processor utilization at that bandwidth. We use the Netperf benchmarking utility to accurately measure bandwidth from a common server to the target client. The server we used was based on the Tyan Thunder K8W which we've shown before to perform at almost full Gigabit speeds.
In the first graph you can see that all three Gigabit chips perform about the same at 100Mb per second which is still the most commonly used network speed. At 1000Mb per second speeds, the Intel chip on the 875P board was a clear leader followed by the Marvell chip on the 925X board. The VIA chip was huge disappoint though as it barely was able to double speed its speed over the 100Mb connection. Processor utilization in these chips were about proportional at 1000Mb speeds. In these chips, faster transfers took more processor usage
Our final graph is not really presented to show you that one platform has a better disk controller, but instead, to show you that they all perform at the same level with the same disk. This shows that disk performance played little to no part in the benchmark results presented above.
