Indeed with the Athlon 64 3800 the framerate improves by 2-3 FPS - empirical testing by playing the game. Probably with the FX-55 you would get ~50 FPS which makes for a pretty good experience (in FEAR it is highly important to have above 40 FPS otherwise the game feels heavy).

Benchmarks: one important criteria is to have some kind of automatic tests even for games so the results could easily be checked by everyone. I dropped all directx 7&8 benchmarks with the sole exception of 3dmark 2001.

Synthetic tests:

DirectX8: 3dmark 2001;
DirectX9: 3dmark 2003, 3dmark 2005, Aquamark3;
GUI test: Tom 2D;
CPU tests: Super pi, Cinebench 2003;
System tests: Pcmark 2002 & 2004 & 2005;

Games:
DirectX9: Farcry, Half life 2; Serious Sam 2, FEAR, Splinter's cell chaos theory;
OpengGl: Doom 3, Quake 4.

I must underline the fact that I have performed all synthetic tests at the default settings first and at maximum details in all subsequent runs. All games have been tested fully maxed out with the exception of Doom 3 and Quake 4 which are too demanding even for GTS 250.
Because I am testing CPU's I do not see the point in testing many resolutions. Most probably 1024*768 would be sufficient, maybe 1280*1024 if there are requests. I dropped 1600*1200 as it is very tough on the video card.

I did tried again Need for speed most wanted but it is impossible to get repeatable results - the framerate varies wildly between the runs even if i choose the same challenge/race so I dropped it from the testing suite.

Now on to testing...

Playing FEAR with both soft shadows and AA 4x is impractical as soft shadows get disabled after restarting the game, even though they show as enabled. This is evident from benchmark results. You don't want to enable it every time you start the game. Therefore I just enable AA 4x.

You want to test CPU in practical scenarios too - 1280x1024 at least. The question is how much newer CPU would help you in practice, not just in low resolution.

Both Quake IV and Doom 3 use the same engine, what about testing Far Cry in OpenGL mode instead (replacing one) ?
Alternative would be Unreal Tournament 2004, which does have OpenGL path, however I'm not sure if usual benchmark tool for it allows it's usage...

1280×1024 added as a second resolution to be tested.
Doom 3 dropped and Farcry tested in opengl.
I am going to specifically test Fear to check if SS are disabled after exiting the game with the GTS 250 and forceware 197.77

While you are testing those CPUs, it would be great to do a boot test with unbuffered ECC memory and registered ECC memory, even at lower speed like DDR 333. Older revisions like Clawhammer may not support it. It would be a test of whether ECC related settings appear in BIOS and whether it boots into Windows XP and ECC shows as enabled in AIDA64. That way people will not waste money on something that doesn't work.

This should probably a separate category of testing.

I still have a s754 NewCastle 2800+ and another Lancaster, no more Clawhammers unfortunately, they got sold. I will test them with unbuffered ECC memory in nForce 3 and VIA K8T800. Maybe I will even get registered ECC very cheap locally.

If someone intends to run memory at 3-3-3-8 they might as well get ECC memory as a curiosity.

I am sorry but I can not accomodate this request now. I have lots of regular DDR all sizes: 128, 256, 512, 1024mb. I got most of them for free. For testing ECC I would have to buy some modules and that means an extra expense. If I manage to get some for free or in a switch, than I will test them asap.

I tried last night FEAR on the current system (Newcastle 3500 + 3gb DDR + GTS 250) and the games remember the settings every single time, including the soft shadows. however, if I go to an earlier driver (possible because G92 is used in many cards) such as 175.19 the game forgets the SS setting and I have to manually enable it every time. So I will test FEAR with SS enabled - GTS 250 is more than capable of mastering this game will all the bells and whistles and 1024*768 and 1280*1024; only at 1600*1200 it drops the FPS significantly.
Clawhammer testing finished and Newcastle is under fire right now. So far the differences in performance are minuscule.

On my PC Soft Shadows reset silently. Game version v1.08.

1.) I follow your guide Re: Battle of the platforms: socket 754! to activate AA 4X + SS. With SS off and AA 4X I get 107 avg fps. When I turn SS on and AA 4X I get 83 avg fps
2.) I leave the game and start it again
3.) I don't change anything, just run the benchmark. This time I get 107 avg fps. SS is clearly off. If I go to graphics quality, it shows as on, but as seen from benchmark it is off.
4.) I go to graphics quality and explicitly turn off SS. I go back to main menu and back to run benchmark. It shows 107 avg fps.

The quickest way to enable SS seems to be:
1.) Disable both AA and SS, go back to main menu
2.) Go to graphics quality and enable SS. Go back to main menu
3.) Go to graphics quality and enable AA 4X. Go back to main menu.

At this point I can run the benchmark and both of them are on. For some reason they conflict and the game doesn't like them being both on so it silently resets SS after restart. This problem exists also in Perseus Mandate.

Does 258.96 driver also has this problem (last driver before OpenGL 4.x support) ?

That remains to be tested. I do not use any driver newer than 197.77 because it causes problem with games like Farcry - terrain looks funny or is missing - and Quak4 4 - textures looks awful. It supports all 6000 series, 7000 series, 8000 series, 9000 series and 200 series cards so is perfect for a XP machine with directx9.

FEAR is also patched up to the latest version with both Extraction point and Perseus mandate installed.

The problem of Soft Shadows resetting is well documented for FEAR. It isn't an issue of driver. It seems the best way to play with both FSAA and Soft Shadows is to enable just Soft Shadows in the game and force FSAA 4X via NVidia driver panel for the app. Unfortunately when I did that, sometimes when benchmark runs there are visual defects (in both FEAR and FEAR Extraction Point). The visual defects look like a shadow/ghost patches on the screen, not color/polygon issues when the GPU is faulty. I have only seen it when FSAA was forced in driver.

Unfortunately the 3500 San Diego that the seller was so sure about it turned out to be a Newcastle! So I am unable to test the last 2 Athlon 64 revisions because San Diego proves to be exceedingly rare and Toledo does not have a 3500 model. However, as you will soon see, I think we can get a pretty good impression of the San Diego performance based on the results you are about to see.

Without further ado, let's begin! The first test is 3dmark 2001, a staple of all retro comparisons. All systems are capable of running any DirectX8 game very well.
The order of the results will always be:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

Second resolution I tested at is 1280*1024:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

3dmark 2001 results shows us that Newcastle is very slightly below the original Clawhammer, Winchester is practically equal and Venice with Manchester provide ever so slightly better performance than Clawhammer.

3dmark 2003 is next. It's the first DirectX9 benchmark and from now on only DirectX9 tests are going to be featured.
First is the default benchmark at 1024*768 without image enhancements:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

Next is 1280*1024 at max details - this is a far more strenuous test than I thought for the Athlon 64 3500! We are heavily CPU bottlenecked!
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

We see a nice progression as we advance through the different Athlon 64 revisions at default settings. At 1280*1024 there is nothing to see as all the results are bunched up closely together.
3dmark 2003 is the only version with 3 sound tests: NO SOUND and 24 SOUNDS are for ""everyone" and 60 sounds are only for hardware solutions such as Creative Audigy. Because AC97 is part of nforce4 this is what I tested.
There is a 5-10% advantage for all 90nm Athlon 64 over the older 130nm Clawhammer and Newcastle. This is repeatable and shows that newer revisions are improved despite the fact that until now we got no difference greater than 5%. Is the first time we see a meaningful difference between different Athlon 64 revisions!

Last 3dmark version to be tested is the 2005 one.
Default settings first:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester