The Gap That Was Not There . . .

A Curious Absence . . .

Up to now, the Athlon 64 has generally competed well against its older brother, the FX. A little bit too well, in my humble opinion.

Someone has finally run the types of tests necessary to demonstrate fairly easily that the 64 does indeed do a little too well compared to the FX.

XBit Labs put up a series of benchmarks testing the A64 3000+.

They ran the 3000+ at its default speed of 2GHz, then overclocked it to 2.2GHz, the same speed as the FX.

This gives you two different matchups to compare. If you compare the scores of the 3000+ at regular speed with that of the 3200+, the only difference between the two is the extra cache the 3200+ has, so any gap in performance will be due to the extra cache.

If you compare the scores of the overclocked 3000+ to the FX, there are now two differences between the two configuration. The FX has more cache, plus it has dual channel memory as opposed to the 64’s single (really single plus 10$) channel memory.

So if we measure the gap in performance between the two 64s, then measure the gap between the 64 and the FX and finally compare the two, the gap between the 64 and FX ought to be generally rather bigger because the FX has an extra performance-enhancing advantage.

If you look back at performance differences between single-channel and dual-channel PIVs in old reviews, you’ll see that the difference is usually a handful of percentage points.

Let’s see what happens when we make that comparison using XBitLabs’ numbers:

I skipped those benchmarks in which the CPU plays only a minor role in differences in performance, or where performance (in games) is obviously video-card limited.

Aquamark3, Default, fps

3000/3200: 1%
OC3000/FX: 1%

Aquamark3, Default, CPU

3000/3200: 1%
OC3000/FX: 3%

Quake3 (four), High Quality, 1024X768X32

3000/3200: 7%
OC3000/FX: 7%

Unreal Tournament 2003 (dm-antalus), 1024X768X32

3000/3200: 4.5%
OC3000/FX: 4.5%

Comanche 4:

3000/3200: 5%

OC3000/FX: 5%

Tomb Raider: The Angel of Darkness (part3) 640X480X32

3000/3200: 4%
OC3000/FX: 4%

X2 – The Threat 1024X768X32

3000/3200: 2%
OC3000/FX: 2%

Business Winstone 2004

3000/3200: 2%
OC3000/FX: 1%

Multimedia Content Creation Winstone 2004

3000/3200: 2%
OC3000/FX: 3%

Data Compression, Winrar 3.2, Best, sec.

3000/3200: 5%

OC3000/FX: 3%

Data Compression, 7zip 3.11, 32Mb, Compressing

3000/3200: 6%
OC3000/FX: 4%

Data Compression, 7zip 3.11, 32Mb, Decompressing

3000/3200: 0%
OC3000/FX: 1%

MPEG-2 Encoding, Canopus ProCoder 1.5, Mastering, sec

3000/3200: 1%

OC3000/FX: 1%

MPEG-4 Encoding, Flask 0.78.39/DivX 5.11, fps

3000/3200: 1%
OC3000/FX: 3%

Windows Media Encoder 9, MP2 to WME, sec

3000/3200: 1%
OC3000/FX: 1%

3ds max 6.0, Final Rendering, Underwater, sec.

3000/3200: 0%
OC3000/FX: 0%

CINEMA 4D, CINEBENCH 2003, Raytracing, CB

3000/3200: 0%
OC3000/FX: -1%

CINEMA 4D, CINEBENCH 2003, Lighting SW, CB

3000/3200: 0%
OC3000/FX: -1%

CINEMA 4D, CINEBENCH 2003, Lighting HW, CB

3000/3200: 3%
OC3000/FX: 7%

Do you see the general pattern? With just a few exceptions, there is no additional gap between the overclocked 3000+ and the FX, especially for games.

It should be noted that the encoding benchmarks aren’t much affected by increased memory bandwidth; they weren’t for the PIV, either.

For games, you’d expect to find the 64/FX gap to usually be 3%-5%-7% bigger than between 64 and 64.

It isn’t.

What Does This Mean?

There are two possible explanations for this:

1) Memory bandwidth doesn’t mean squat when you have a low latency platform or

2) The FX is running on a platform (socket 940) that is meant for servers and is not tweaked for desktop speed, while the socket 754 platform is.

Item one may be a factor, but I would bet my money that the major reason why the gap that should be there isn’t there is because of door number two.

So the reason why the 64s look so good against the FXs is because the socket 940 platform is a bit of a slouch running desktop benchmarks and games. This is very understandable given the nature of the platform. To cite just one reason for this, socket 940 uses registered RAM. To make a long story short, registered RAM sacrifices a little speed for more stability.

Yes, of course it does very well, but it could do even better with a configuration more optimized for desktop apps and games.

And it will.

What does this all mean?

That means when the socket 939 boards that are meant for the desktop come out, the gap that isn’t there now will reappear. For things like gaming, it very roughly will mean that a socket 939 Newcastle ought to do about 5% better than a socket 754 Newcastle running at the same speed.

That all by itself isn’t a killer factor to wait, but when you add that to the fact that socket 939 is going to become the mainstream Hammer platform and socket 754 will be orphaned out to the value processor home, it’s an additional good reason for most to wait for just three more months.

There’s a lot of people saying to buy the 3000+ right away, and only a few making the points I’ve been making. Three months from now, the same people who are telling you to buy socket 754 today will be telling you to buy socket 939 tomorrow.

Come to think of it, how often do those folks tell you NOT to buy whatever they’re reviewing at that moment? Think about it.

Don’t get manipulated and bullrushed into buying prematurely. Socket 939 will be the better buy.

Ed