MaximumPC 2004 09

CPU Athlon 64 3400+ Athlon 64 FX-51 % Difference Athlon 64 FX-51 % Difference Athlon 64 3400+

Clock speed 2.2GHz 2.2GHz 2.2GHz 2.2GHz

Socket 754 pin 940 pin 940 pin 754 pin

Mainboard Soyo CK8 Asus SK8N Asus SK8N MSI K8N Neo

Chipset nForce3 150 nForce3 150 nForce3 150 nForce3 250Gb

RAM 1GB DDR400 single

channel

1GB DDR400 registered

single channel

1GB DDR400 registered

dual channel

1GB DDR400 single

channel

Quake III “Normal” Four 450 435 3.3 450 0 456

Sandra RAM Composite 3041 3034 0 5609 84.5 3005

3DMark 2001 SE 19990 19511 2.4 19820 -2.4 20095

AquaMark 3 CPU 9638 9243 4.1 9829 2 9665

3DMark 2003 CPU 757 729 3.7 780 3.7 757

UT2003 Fly By 6x4 309.3 305.1 1.3 312.3 .9 317.0

SYSmark 2004 Overall 173 169 2.3 174 .6 168

T

he debut of AMD’s Socket 939

signals a rapid sunset on the older

Socket 940 line. In a nutshell,

Socket 939 and the Athlon 64 FX

processors that use it eliminate the

need to use registered RAM modules.

Registered RAM uses the same basic

SDRAM chips as today’s DDR400

modules, but includes an additional

controller chip that “redrives” the

signals to the memory chips for

more reliable communications. This

reliability isn’t free, of course. The cost

of registered RAM is an additional clock

cycle of latency.

Registered RAM is mostly used

to increase reliability in servers and

workstations, which are densely packed

with multiple gigabytes of memory.

Desktop machines, which aren’t as

prone to 24/7 operation or housing large

amounts of RAM, usually don’t need

nor even work with registered RAM.

But in creating the fi rst Athlon FX for

Socket 940 boards, AMD essentially

changed the name of its workstation/

server CPU from Opteron to Athlon 64

FX and carried over the registered RAM

requirement.

To gauge the performance impact

of migrating from registered RAM to

normal RAM, we took our reference

Socket 940 Asus SK8N motherboard

and outfi tted it with 1GB of registered

RAM and an Athlon 64 FX-51. We ran

this board in both single-channel mode

and dual-channel mode with the same

RAM. We then confi gured a Socket 939

Soyo CK8 motherboard with 1GB of

DDR400 and an Athlon 64 3400+. Both

boards use the nForce3 150 chipset, as

well as the same hard drive, videocard,

driver, and OS, making the test a fair

measurement. While the design of the

motherboard and BIOS could have

a slight impact on our results, it’s

the fairest way to isolate the variable

of different kinds of memory. And

although it can’t be directly compared,

we also confi gured a machine using an

Athlon 64 3400+ in an MSI K8N Neo

board equipped with the newer nForce3

250Gb chipset. All the systems’ CPUs

were clocked at 2.2GHz.

Our Lab tests yielded interesting results.

First the most obvious: Registered RAM

could not have been removed from the

Athlon 64 FX soon enough. It’s clear

that the extra latency has handcuffed

the Athlon 64 FX line. The most direct

comparison was between the SK8N/Socket

940 board running in single-channel mode

and the CK8/Socket 754 in single-channel

mode (Socket 754 CPUs do not support

dual-channel modes). The registered RAM

system trailed the non-registered RAM

system by margins up to 4.1 percent.

Also of interest is how well the Athlon

64 FX-51 performed when running in

dual mode—while it was faster than the

non-registered system, it won by less than

we expected. We saw about a 3.7 percent

bump in 3DMark 2003’s CPU test and

very slight increases elsewhere. Even more

surprising, the dual-mode FX system and

the single-channel system tied in Quake

III. And weirder still, in 3DMark 2001 SE ,

the FX ran behind the non-registered

system. The only improvement we saw

was a massive 84.5 percent bump in the

synthetic SiSoft Sandra RAM test. This

leads us to believe that either synthetic

benchmarks aren’t worth a damn or

that there are very few applications that

can actually use or are optimized for the

boatloads of memory bandwidth the

Athlon 64 FX provides.

In the Lab A behind-the-scenes look at Maximum PC testing

The Cost of

Registered RAM

We test how much faster non-registered

RAM runs

 MA XIMUMPC SEPTEMBER 2004

Registered RAM, like this Corsair Micro

module, isn’t necessary in the new

Athlon 64 FX PCs. Old FX-51 CPUs

required it, but new versions don’t.

normal RAM, we took our reference Registered RAM, like this Corsair Micro