And 1280*1024 maximum settings:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

Nothing to see here, we need a stronger CPU and a stronger video card! Moving right along to...
Aquamark3, the very first DirectX9 benchmark ever to be released. In order to keep the topic clean and tidy I will post only the screenshots with the average FPS and not the screenshots with the details FPS for each test.

I’ve got a socket 939 i never use all the caps are leaked.
did some research and found they were manufactured in capacitor plague times.
Its a Gigabyte and had one owner from new I am second owner. It works surprisingly.
I have little confidence with 939 for several reasons.
Bugs were produced with the chipsets and memory control most popular is its inability to reach full advertised memory capacity.
Since 939 chipsets were the same on all vendors this is important.
I recommend anyone please do research on these if your thinking retro building 939.
The information about them is legendary multiple sources.
I having the hardware people threw away 939 for its unlikely I’ll ever refurbish this gigabyte lemon of mine.
Sometimes I just look at it wondering how they got away with their lemon marketing.

San Diego with 1MB L2 cache is needed. We see performance bump from Venice, but with 1MB L2 cache it would be even greater. This is why Lancaster in s754 does well.

3d mark 2003 sound test seems to be memory bandwidth sensitive. So in real world bandwidth matters, whereas synthetic benchmarks or even game benchmarks don't show it sufficiently as not all play sound.

Socket 939 is indeed legendary! AMD overthrew Intel and took the performance crown and held it for 2 years!
1. capacitors are a problem with certain brands/models -maybe your Gigabyte is one of them - all Abit boards have Rubycon capacitors that are in great shape. This was a problem affecting both AMD and Intel boards at the time.
2. Chipsets are not the same, in fact I demonstrated that performance differences remain despite the Athlon 64 having an integrated memory controller.
3. They reach full memory capacity - you just have to use DDR-333 with early Athlon 64's - but you do not need actually 4gb as XP sees only 3.25GB. I already explained how you can have 4 memory modules running at DDR-400 with all Athlon 64 revisions.
4. After around 500 tests my personal experience with both socket 754 and 939 until now is: they have some quirks, but nothing that I could not solved. Stability and performance is as people remember it: extraordinary! Of course there are some platforms that stand up: nforce4 is simply outstanding until now.
More tests remain to be done before making a definitive conclusion.

AlexZ wrote on 2026-08-04, 08:56:

San Diego with 1MB L2 cache is needed. We see performance bump from Venice, but with 1MB L2 cache it would be even greater. This is why Lancaster in s754 does well.

3d mark 2003 sound test seems to be memory bandwidth sensitive. So in real world bandwidth matters, whereas synthetic benchmarks or even game benchmarks don't show it sufficiently as not all play sound.

I got Newcastle and Venice 3800 -2.4 GHz - and compared them in 3dmark against Newcastle and Venice 3500. They are around 5-6% better - so 200 PR points. Taking into account that Athlon 64 4000 has 1MB of L2 cache (that equals 200 PR points) it should be 10-12% better than 3500 and it should be a match for the GTS 250 when playing any game up to 2005 at max details.

Aquamark3 at default settings puts up a score that we are not interested in. The average FPS is what we are looking for.
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

Because Aquamark3 is the second synthetic benchmark that has sound as an option I also tested with audio enabled:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

Aquamark3 poses no problem for the GTS250 which is strongly limited by the CPU!
1280*1024 is next. Without sound first:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

... and with sound:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

In Aquamark3 Newcastle is again last, Clawhammer just in front of it, while Winchester, Venice and Manchester each provide - successively - a very small boost. Again in the benchmarks that include sound the 90nm Athlon's manage to distance themselves from the 130nm ones! This is the second time where there is a 5% difference between the CPU's tested - so beyond the margin of error.

Synthetic 3d tests are done, time to move on to overall system benchmarks with PCMARK 2002, the oldest version available. The memory controller inside the Athlon 64 improved with each generation so let's see if we can spot some differences!
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

Just as predicted, Winchester, Venice and Manchester have 5% better memory scores than the older Clawhammer and Newcastle! CPU scores also improve with a measly 3% on Venice and Manchester but that is well within the margin of error.
Next is PCMARK 2004, a completely overhaul suite that add a graphic score.

PCMARK 2004:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

Newcastle performance is seriously underwhelming. Maybe cost cutting did not constitute such a good idea, after all! Other than that all other revisions of Athlon 64 perform within 1% of each other!

PCMARK 2005 is the last version compatible with Windows XP.
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice

5. Manchester - where I forgot to check the graphics test the first time I ran it.

All CPU and memory scores are within 2% so no differences here. What strikes immediately is the poor graphics score of the Newcastle and Winchester Athlon's. Maybe games will suffer as a consequence.
We shall see when we reach the games.

Next test is SuperPi - a processor benchmark that is single threaded; next time I am going to replace it with something multi-threaded:

1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

Nothing to see here -All Athlon 64 perform the same.

We advance to the next CPU test: Cinebench 2003 that is going to be replaced with R9.5 or R10 next time:
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester

From Clawhammer to Manchester all Athlon 64 perform the same.

And now the most controversial test: Tom2D! You need a specific combination of hardware and software in order to make Tom2D "work".
1. Clawhammer
2. Newcastle
3. Winchester
4. Venice
5. Manchester
Unfortunately all CPU's have serious problems with GUI acceleration in Windows XP!