Navigate: Advice Beginners BIOS Guide CPUs Links Mainboards Memory Network Storage Video/Sound Cards Contact Forum SiteMap Sponsors WebNews Home	.	.
Prices: Mainboards ABIT ASUS Chaintech Shuttle Soyo Tyan CPU Intel P4 2.4C-800 P4 2.6C-800 P4 2.8C-800 P4 3.0-800 P4 3.2-800 AMD AthlonXP XP 1700+ XP 2000+ XP 2400+ XP 2500+ XP 2700+ XP 3000+ XP 3200+ Athlon64 Athlon64 3200+ Athlon64 FX-51 Opteron Opteron 240 Opteron 242 Opteron 244 Opteron 246 Memory Corsair Crucial Kingston Mushkin OCZ Search Prices:

LOSTCIRCUITS	SHORTCUTS: The Missing Pin The Numbers Game / Test Configurations Latencies Memory Bandwidty Sysmark2004 Sysmark Continued 3ds max, Amorphium3 Cinebench3 Gaming Benchmarks More Gaming Benchmarks Overclocking, Conclusions Give Us Some Feedback to Help Us Improve our Reviews
AMD Athlon64 FX53 Back to the top
(Review by MS, July 1, 2004)

Summary

AMD has finally released the Socket 939 version of their Athlon64 processors, which includes the Athlon64 FX53 (Socket 939), the Athlon64 3800+ and the Athlon64 3500+ with the two latter running the dual memory controller but being restricted to the smaller 512 kB L2 cache of the Newcastle die. However, the entire family of Socket 939 processors is capable of running unbuffered memory, which shaves off one bus cycle on each initial memory access.

Aside from the price advantage over the Socket 940 processors, is there any gain in performance caused by the lower latency memory access? Moreover, does the cache reduction hurt the performance and if so, what are the critical applications?

We have looked at the performance of the new platform from a variety of different angles, including different memory configurations that can make quite a difference. After all, AMD processors are more dependent on low latency memory and, by extension, on open pages..

AMD’s Athlon64 line of processor has become almost an instant repetition of the original Athlon with respect to becoming an instant success. Times have changed, it is no longer the Pentium3 and, er, ar, ..DRAM that is the enemy, rather it is the Pentium4 and as of latest, DDR2. We’ll save that, however, for a different story. Despite the fact that the transition from the 32-bit environment to a fully functional version of Windows still substantially lags behind the hardware, even the execution of 32-bit software has set new standards for many applications. Needless to say that this is true not only for a Windows-On-Windows environment on a 64-bit operating system but that it equally holds for the plain vanilla 32-bit version of WindowsXP.

The family of AMD Athlon64 processors is quite diversified and a correlation of the different features with the different product monikers is more than just slightly counterintuitive. Briefly, there are two versions of the core that are different with respect to the amount of integrated cache, since the so-called “Newcastle” core only features 512 kB of L2 cache whereas the original Clawhammer core features the full 1 MB. Both versions are available in Socekt 754 format. Additionally, the AMD repertoire features the Sledgehammer core originally released as Opteron 14X with its 940 pin Socket interface, featuring dual memory controllers, whereas, the more plebeian versions fitting the Socket 754 only sport a single memory controller. Adding extra confusion to the whole mess is the fact that AMD model numbers are floating, that is, for example, the original Athlon64 3200+ was running at 2000 MHz, featuring the full 1 MB L2 cache, the new ones are running at 2200 MHz but lost ½ of the cache, which is, therefore only 512 kB. Winchester and San Diego cores will be manufactured using the 90 nm interconnect and replace the current Sledgehammer and Newcastle in the Athlon64 and the FX versions, respectively.

      

Click for larger picture.

The latest differentiation occurs on the level of the pin-out and reflects the differences in the requirements for the type of memory used. That is, the original Sledgehammer / Opteron / Athlon64 FX with its socket 940 pinout required the use of registered DIMMs and would not work with standard “unbuffered” or unregistered modules. Briefly, the difference is the presence of an interface chip, which is the only IC that the memory controller actually sees.

The very nature of registered DIMMs and their interface chip implies that the latter will not be able to propagate the addresses and commands to the discreet memory chips on the fly, rather, in the interest of precise timings, one penalty cycle is inserted, and the signals are then forwarded on the next rising clock edge. It is understood, therefore that on every initial access one additional memory latency is incurred, courtesy of the register. This is where the socket 939 makes a difference in that it supposedly combines the best of both worlds, namely the dual channel with the capability to use unbuffered memory. By extension, that means that the high bandwidth of the dual channel meets the low latency of the standard controller and the result should be an unparalleled data supply to the processor. The only drawback is that the Athlon64 3800+ only has the small “Newcastle” Level2 cache of 512 kB.

Otherwise, there is relatively little to say about the Socket939 other than that this is what everybody has been waiting for since the original release of the Athlon64 --- minus the ½ cache. --- It took awhile to see the originally projected product and now the question is whether it was actually worth the wait or whether time has passed this particular processor by. After all, other than in synthetic benchmarks, it appears as if the much simpler Athlon64 with the single memory interface has found its own clansmen who are interested more in the price/performance ratio and the built-in upgrade path to a 64-bit OS than in just performance at a price, -- regardless of whatever will be in the future..

next page: => Current Models and Numbers -- Test Configurations =>