[an error occurred while processing this directive]

Five AMD Athlon 64 Models – Which One to Choose?


RightMark Memory Analyzer

Let's start the tests with benchmarking their operating speed with memory in RMMA. These synthetic tests for processors will help to make sure that all the systems operate in their normal modes without performance loss.

This test is necessary because the first CG stepping tests of processors demonstrated unexpectedly low results on some motherboards. Having thoroughly analyzed the situation, we have come to a conclusion that not all BIOS versions are identically effective. The fact is that processors of a new stepping (unlike the previous C0) can operate with more than two memory modules at 400 MHz. However it requires the special "2T Command Rate" mode, which improves operating stability with a large number of banks, however it considerably slows down memory write speed: the tests demonstrated only about 2600 MB/sec instead of standard 3200 MB/sec. Surely this reduction resulted in poor test results in real applications.

The problem turned out in BIOS versions used in tested motherboards. Though only two modules were installed and optimized settings were used, 2T Command Rate parameter would turn out enabled. Sometimes the problem was solved by forcing it disabled (it does not have effect on PC stability), and in the hardest cases we had to disable it using the wpcredit program. At the same time we have discovered another configuration trick for the integrated controller in K8 series processors. When using more than one memory bank (there are usually two, four, or eight banks), you can use the interleaving mode, which may add a couple percents of performance gain. Some BIOS versions enable this mode automatically, the other ones don't even have a corresponding option in BIOS SETUP. Unfortunately, this mode cannot be enabled on the run, but with some skill you can always check its status (for example, using the wpcredit).

The first diagram – memory read speed. We can see a distinctive division of AMD products by the integrated memory controller into single channel and dual channel supporting devices. Note that the measured read speed was close to maximum.

Interestingly, Athlon 64 FX-53 equipped with registered memory modules has a higher read speed (by 1.6%) than the new model for Socket 939, operating with regular modules. On the other hand, the difference is too small to pay any special attention to it. Let's see what's going on in the other tests.

The second diagram shows the write speed, it is practically a copy of the first one. Again AMD products are divided into two groups. This time the new Athlon 64 FX-53 is not lucky again – it lags behind by 1.68% from its predecessor.

The third and fourth diagrams demonstrate the latency of memory operations for tested processors. At last we see the considerable difference between models with registered and non-registered memory this time. Latencies are much higher for Athlon 64 FX-53/940. Let's see what effect it will have on tests in real applications, to which we now proceed.

* the results for Pentium 4 CPUs are obtained in RMMA 3.2 according to the method described in this article.

3ds max

Athlon 64 FX-53 is the leader among AMD64 processors, note that it uses the old socket. The other results demonstrate that the performance in 3dsmax mostly depend on the CPU clock (in case of the same architecture), while the two channel mode and L2 cache capacity do not play an important role.


The result is even more illustrative than for 3dsmax – frequency, frequency, and frequency again. So if you buy a processor to work in these programs, you can save a decent sum of money: Athlon 64 3400+ is almost three times as cheap as Athlon 64 FX-53.


That's the third application that we see "interested" only in frequency. All AMD64 processors with the same frequency demonstrate absolutely identical results.


Ah... it's not clear how to comment on it. Frequency is again the determining factor in this test. Though it's slightly dependent on operating speed with memory (single channel solutions lose about 3%).

Windows Media Video 9

Let's go on... "The picture is absolutely identical to the previous" (C).

Canopus ProCoder

This program is a little more interesting – you can see a strong dependence of the encoding speed on L2 cache capacity: Athlon 64 FX-53 is almost 6% faster than Athlon 64 3800+.

Mainconcept MPEG Encoder

Counting applications critical only to the core frequency (certainly with a reservation "for one architecture"). That's six...


In the (Win)RAR test the fastest processors are the ones with large cache (transition from 512 KB to 1 MB "costs" 6% of speed). And as always, the frequency has a very noticeable effect. Interestingly, processors with the same 3400+ rating produce the same results here.


Results of this test are appealing – there is not a single pair of similar results, except for Athlon 64 FX-53 for 940 and 939 sockets. One can say that 7zip wants everything – frequency, larger cache, and dual-channel access to RAM. A maximum spread of results among 2.4 GHz processors is over 15%.


And finally we see a performance gain due to abandoning registered modules in Athlon 64 FX-53! Of course, 2–3% is not much according to modern standards, but it's nice when practice accords with theory at least a little.

Besides, the games profit considerably from using dual channel memory controllers in Athlon 64 FX-53 and Athlon 64 3800+. Performance in games is less dependent on L2 cache. As in the Win(RAR) test, processors rated 3400+ have the same results in Serious Sam 2.


Kirill Kochetkov (kochet@ixbt.com)
Dmitry Majorov (destrax@ixbt.com)

December 9, 2004

[an error occurred while processing this directive]