agent_x007 wrote on 2025-08-01, 16:50:

old school gamer man wrote on 2025-08-01, 16:27:

agent_x007 wrote on 2025-08-01, 16:18:

Warning to anyone thinking about this combo : Those three things do not work together by default.
A driver hex edit hack is required, which allows 1440p60 mode to be available (default is 1080p60).
Unless you have VGA capable 1440p monitor, then you are good to go 😀

Is this only a problem with the 750 in xp? I had a hand full of cards that all worked with both 1440p and 1536p in xp out of the box if not by adding it in the nv control panel.

It's HDMI issue, but DP is also limited to HBR1 speed (so 1440p at ~90Hz max.).
Some context : Re: Windows XP at high resolutions with scaling

hmm odd I never had that problem. maybe because I was not using 90hz ?

AlexZ wrote on 2025-07-29, 23:46:

I have tried to stabilize 4x 2GB at 1066 but it wasn't stable even at 2.3V. Cas latency is 9.375 ns with 5-5-5-15 which is probably too short for 4 modules.

I didn't read the entire thread, but why are you using 4x 2GB mem sticks, and not 2x 4GB? They do exist in DDR2 format.

pixel_workbench wrote on 2025-08-02, 20:49:

I didn't read the entire thread, but why are you using 4x 2GB mem sticks, and not 2x 4GB? They do exist in DDR2 format.

4GB DDR2 modules are very rare, I didn't find any locally. The motherboard claims to support 16GB memory.

A fun system and a few benchmarks:

System:

3dmark2003:

3dmark2005:

Crysis:

And the temperatures after all the benchmarks:

Was finally able to get myself a HD3870x2 which is not dead. After cleaning, replacing thermal compound and swapping second GPU heatsink for copper one from another card i decided that AM2 would be a good platform for this card to run a few benchmarks.

Interestingly enough temperatures are good (this uses AMD box aluminum cooler with 60mm fan too), though it does sound like a vacuum cleaner under load.

Performance is... interesting. It is definitely not bad, but not amazing either. Also compatibility with games is even worse than SLI - even crysis had some artifacts like flickering textures.

In synthetic benchmarks it isn't bad at all, but a total flop in crisis which makes it useless for AM2 unless there is a driver problem that can be fixed. At DDR2 750 your timings should be 5 5 5 18.

I finally put together my AM2+ build with Phenom II X4 955 BE 3.2Ghz and PNY GeForce GTX 780 and old friend coil whine is back just like in my Phenom II X6 build. Different board, different GPU, different PSU yet same result. This is one of advantages of Athlon 64 3400+. There is no coil whine as the load is so small.

AlexZ wrote on 2025-08-29, 20:34:

In synthetic benchmarks it isn't bad at all, but a total flop in crisis which makes it useless for AM2 unless there is a driver problem that can be fixed. At DDR2 750 your timings should be 5 5 5 18.

Yep, this is how it goes with multi-GPU setups. Synthetics work, everything else - sometimes. I bet if i spent some time looking for games which actually work well with this card the results would be quite good, but crysis definitely did not like it. Driver is 13.4 which worked well with 3850/3870 before...

This was even evident in some period-correct reviews - some games were like 20% better than the competition, some - 50% worse. But otherwise feels a little bit like a scam - you look at nice numbers in synthetics and specifically picked games, buy the card and get... not what you expected at all.

One GPU here is also a bit slower than single 3870 because it uses GDDR3 instead of GDDR4.

I am glad multi-GPU in this form is no longer a thing...

Memory... yeah, this are completely default settings. I am sure if i tweaked it it would have worked. But i did not consider it important for the purpose of this test and this board probably should not be used with 125W CPU at all so this is by no means a permanent setup...

What's curious - this board being only slightly newer than S939 stuff and completely low end - it has working and configurable fan control. Not user-friendly at all as a bunch of non-obvious numbers have to be configured, but whole lot netter than those fancy asus S939 board...

AlexZ wrote on 2025-08-29, 20:34:

I finally put together my AM2+ build with Phenom II X4 955 BE 3.2Ghz and PNY GeForce GTX 780 and old friend coil whine is back just like in my Phenom II X6 build. Different board, different GPU, different PSU yet same result. This is one of advantages of Athlon 64 3400+. There is no coil whine as the load is so small.

Yeah... coil whine. Have not encountered it for a while, at least in severe enough form to be annoying. Intel ARK750 i bought for kids to play modern games some time ago does have it, but not loud enough to really matter.

If it is from the motherboard sometimes disabling all the power saving stuff can change or remove it completely. If it is from GPU... basically tough luck. I remember having GTX470 back then which suffered from it...

So, i remembered that on all this CPUs multiplier is only limited by maximum value and can be freely set lower. And thought it would be fun to compare AM2 with S939 to see how much the performance increase actually was at the same frequency. It turns out the BIOS on ECS C51GM-M is absolutely brilliant, better than on both S939 boards i have and also fancy abit AM2 board. There is everything including CPU-NB, NB-SB, memory and CPU multipliers, voltages, memory timings, etc... So i've set multiplier to 12 and got pretty much exact equivalent of S939 4800+:

And then i ran some benchmarks:

3dmark2001:

3dmark2003:

3dmark2005:

Aida64 CPU/memory benchmarks and crysis:

Everything at 1024x768 with 8800GT.

Very curious results. The difference is there, probably because of faster memory, but it is very small. So small that it would make switching from S939 to AM2 pointless back in the day, unless the CPU itself was significantly faster, like 6000+. Also worth remembering that back in early AM2 times DDR2-800 was kind of pricey, judging by discussions i was able to find it could be as much as 2x more expensive than DDR1-400 or DDR2-533.

And that is the best option for AM2 - windsor with 2x1MB cache. Would be quite curious to see how brisbane works - will it be really slower than S939 equivalent? By the looks of it certainly not going to be faster... Is it really possible that 2-3 year newer CPU is slower? Weird times for AMD and it kind of shows where things went wrong. S939 was good, better than first intel netburst dual cores. But then intel made Core2 while AMD was just rebranding and re-releasing the same core for 3 years with die shrink which caused performance regression...

Also not sure why in some benchmarks it feels like GPU is slightly faster on S939. It is not. Exactly the same card. Something else must be causing the difference...

Your benchmarks demonstrate there is little performance difference between Athlon 64 X2 4800+ on s939 and AM2. We can think of Toledo and Windsor as basically the same product, but Windsor having DDR2 controller built-in and manufactured on refined 90nm process so it can reach higher clocks. Then came Brisbane and Phenom 1 and things deteriorated. AMD either hoped 65nm would bring higher clocks or considered L3 cache but had to skip it until Phenom. Brisbane will be probably in the same league as s939 with 512kb L2 cache. With AM2 you can quite accurately simulate s939.

Note how memory performance improves with CPU clock speed. This is also noticeable in Everest and Sisoft Sandra. At 2.2-2.4Ghz there was little benefit from DDR2 667. Windsor at that clock speed is unable to utilize DDR2 bandwidth. Things only get better above 3Ghz provided you tune memory timings.

More notable benefit of AM2 was higher memory capacity of DDR2. You could use up to 8GB RAM. It is an advantage in Windows Vista. Another benefit was black editions of CPUs.

AMD wasn't that terrible, but Intel simply had better products, being one steap ahead. AMD's answer to Intel Core architecture was Phenom, which was a failure at first.

s939 covers Windows XP era (2002-2006) and fails in Windows Vista era (2007-2009) due to low clocks. AM2 can cover also Windows Vista era provided you do not have too high expectations.

Socket 939 had access to Winbond BH memory which wasn't topped out by DDR2 for quite a while. Only 2 Gb possible though. Also Athlon 64 overall was responsive to low timings, which was not a thing for DDR2.

Another benefit was black editions of CPUs.

Black Edition wasn't a benefit in any form, because 90nm Black Edition CPUs were locked and 65nm had shit overclocking anyway.

DDR2 800 can do 4 4 4 12 at 2V. I have not seen DDR 400 memory on sale locally that could do 2 2 2 6. I would have bought it and put it into my Athlon 64 s754.

AM2 Athlon 64 X2 6400+ BE (ADX6400IAA6CZ, ADX6400CZWOF) is unlocked. It is the only black edition Windsor.

Like I said, Winbond BH. That memory was capable of achieving 500Mhz with 2-2-2-5.

AM2 Athlon 64 X2 6400+ BE (ADX6400IAA6CZ, ADX6400CZWOF) is unlocked.

It's not unlocked, it has the highest default multiplier and that's it. FX series allowed multipliers way above 16. Also on majority of motherboards adjusting multiplier in any way led to broken energy saving.

I have not messed with timings (and memory in general) here on purpose, leaving everything at defaults. On S939 the memory i have refuses to go lower. Even 2.5-3-3-8 causes errors already, that i tested. Even though the same memory works much better on S462. And i do not want to overvolt given rarity of this CPUs.

On AM2 i could have probably done better but then it just adds more variables into comparison...

AlexZ wrote on Yesterday, 17:55:

s939 covers Windows XP era (2002-2006) and fails in Windows Vista era (2007-2009) due to low clocks. AM2 can cover also Windows Vista era provided you do not have too high expectations.

What's curious to me - in low-mid range AM2 is not an improvement over S939 at all. Back in the day i had 4200+ 512k windsor or something like that. Turns out it was no better than 4200+ manchester on S939.

Only top end 90nm AM2 CPUs, like ~5600-6400+ with 2x1MB cache would have been a noticeable improvement and those were later addition and were not manufactured/sold for long AFAIK, being replaced by inferior (but cheaper) brisbane.

Back then this was very good level of performance, not "the best" possible, but good. For Vista and even 7 in the beginning. Nowadays... honestly from what i've seen so far i would consider all this CPUs only good for XP for the purpose of running games. Yes, vista and 7 stuff can run, but this CPUs can not offer "flawless" performance and i see very little reason to settle for less...

AlexZ wrote on Yesterday, 17:55:

AMD wasn't that terrible, but Intel simply had better products, being one steap ahead. AMD's answer to Intel Core architecture was Phenom, which was a failure at first.

AMD was ahead all the way to core2 release, Athlon64 was great when compared to netburst, dual or single core. That's partly why S939 is fascinating to me.

That ended with core2 which coincided with AM2, though AM2 still was not horrible. Then things only got worse...

AlexZ wrote on Yesterday, 17:55:

More notable benefit of AM2 was higher memory capacity of DDR2. You could use up to 8GB RAM. It is an advantage in Windows Vista. Another benefit was black editions of CPUs.

Well, not on a board with 2 slots 😀

And technically it is possible to run 8GB of registered memory on S939, something i am very curious to try and see how much it affects performance.

Also nobody really considered so much memory back then, it would be kind of like buying modern system with 128GB. Possible, but pointless. From what i've seen digging through old forums people were seriously considering high-end multi thousand $ builds with RAM configurations like 2x512MB. Yes, even for vista. It was not obvious right away just how much memory it needs and this is probably the main reason for its poor reputation...

By the way, I just tested 2 versions of the Athlon64 3800 for socket AM2, revisions F and G. Well, the G revision is so much slower, it feels like using some Pentium 4.

That G revision was probably Brisbane, it had G1 and G2. I have tested it, see Re: Any love for AM2? . Brisbane is only worth it for collection if you can get it very cheap, but not for retro builds.

And technically it is possible to run 8GB of registered memory on S939, something i am very curious to try and see how much it affects performance.

Well, memory overclocking will be crap, also this will work only with Command Rate 2T, while Registered Memory internally slows it down to 3T, I think.

pixel_workbench wrote on Yesterday, 19:27:

By the way, I just tested 2 versions of the Athlon64 3800 for socket AM2, revisions F and G. Well, the G revision is so much slower, it feels like using some Pentium 4.

Curious. G should be brisbane and it is slower, but it should not be significantly slower. Also 2Ghz brisbane is 3600+, unlike 2Ghz windsor (F), which is 3800+. I do not think there is a brisbane which is marked as 3800+.

The difference should be pretty small anyway, definitely should not feel like P4.

Unless... are you talking about single core ones? Have never considered single core for AM2. Those should be pretty much the same as dual cores but with different names. F is 90nm G is 65nm...

The Serpent Rider wrote on Yesterday, 19:47:

Well, memory overclocking will be crap, also this will work only with Command Rate 2T, while Registered Memory internally slows it down to 3T, I think.

Yep, still curious how it would affect performance. And how 2x2GB registered would compare to 4x1GB. With 4 sticks 2T has to be used regardless and that is one of the reasons i am using 2x1GB for now - i've noticed measurable performance difference with 1T vs 2T even given the rest of the timings are the same.

Actually, the single core ones are the surprisingly slow ones. I also tested a A64 x2 5000 BE, which is a dual core Brisbane, and it performed close enough to where I expected.

But the single core rev G (Lima) was a lot slower than rev F (Orleans) . Both were 2.4ghz A64 3800, but where one would get about 50 fps in a game, the other ran in the 30s.

According to Everest latency tester, the single core rev G has somewhat slower L2 cache, but the memory latency is a lot worse, like 130ns. I almost suspected a bad CPU, but it passed all my stability tests

Archer57 wrote on Yesterday, 19:23:

And technically it is possible to run 8GB of registered memory on S939, something i am very curious to try and see how much it affects performance.

Also nobody really considered so much memory back then, it would be kind of like buying modern system with 128GB. Possible, but pointless. From what i've seen digging through old forums people were seriously considering high-end multi thousand $ builds with RAM configurations like 2x512MB. Yes, even for vista. It was not obvious right away just how much memory it needs and this is probably the main reason for its poor reputation...

I agree with your assessment on the AM2. In hindsight it doesn’t seem like enough of an upgrade to deserve a new socket. DDR1 was lower latency which was good for gamers. You had to ramp up the DDR2 memory clock speed on AM2 just to break even with 939. AM2 seems more like a preemptive cost cutting move. So many AM2 boards seemed low quality. My socket 939 opteron 175 lasted a long time and it wasn’t until the E8400 came along that I found a compelling upgrade. This board was a powerhouse for years for me:
https://theretroweb.com/motherboards/s/asus-a8n-sli-deluxe

I disagree about the contemporary RAM usage though. There are niches for developers that use a lot of RAM. They spin up a lot of VM’s & containers on their desktops these days. But for gamers, yeah, no benefit.

The Serpent Rider wrote on Yesterday, 19:47:

And technically it is possible to run 8GB of registered memory on S939, something i am very curious to try and see how much it affects performance.

Well, memory overclocking will be crap, also this will work only with Command Rate 2T, while Registered Memory internally slows it down to 3T, I think.

It's fine.

Just to put things into context, here are Aida64 results for my s754 Athlon 64 3400+ and AM2+ Phenom II x4 955. Athlon 64 3400+ memory timings haven't been tuned and it runs on auto and 2T. Phenom II x4 955 doesn't work with 1T command rate.

What is interesting is how slow L2 cache was in Clawhammer. DDR2 gives us bandwidth not far away from that L2 cache except for latency.

L2 cache in Phenom II is really fast compared to s754 and s939. L3 cache is faster than L2 cache in Clawhammer.

It would be great if Archer57 could post his results for Windsor at 3Ghz.

DDR timings tuning guide https://www.elektroda.com/rtvforum/topic1203122.html