AlexZ wrote on 2026-04-15, 09:22:

We will see whether it works at 2.4Ghz without voltage adjustment.

It depends on the operating system.
For example, my Turion runs perfectly at 2.2 GHz on its native voltage, which the motherboard sets correctly.
But as soon as the OS driver loads (whether the one built into Windows XP or the separate AMD Cool-n-Quiet driver for Windows 2000), the voltage is set to 1.45V.
Just like a standard Venice, not even 1.3V like the low-power Venice.

That is weird because I do have the AMD Cool&Quiet driver installed in Windows XP (power profile set to always on) and voltage is set to 1.4V, see screenshots. The driver doesn't cause any problems. I didn't try Linux to see what voltage would be reported there. Gigabyte BIOS just reports OK voltage.

AlexZ wrote on 2026-04-16, 07:34:

That is weird because I do have the AMD Cool&Quiet driver installed in Windows XP (power profile set to always on) and voltage is set to 1.4V, see screenshots. The driver doesn't cause any problems. I didn't try Linux to see what voltage would be reported there. Gigabyte BIOS just reports OK voltage.

If I’m not mistaken, my Turion runs at 1.2V.
But 1.35V here https://www.cpu-world.com/CPUs/K8/AMD-Turion% … ML30BKX5LD.html
It doesn’t really matter, even at 1.35V – the difference is still two voltage adjustment steps for the Athlon 64.
Similarly, I can’t check this in the BIOS either, because it says ‘Vcore – OK’ ~wbr Gigabyte.
But I’m absolutely certain about how Windows 2000 behaves, because I was unpleasantly surprised by the increased noise and heat after installing the AMD Cool’n’Quiet driver.

So there were indeed no fun and no reasons in ocing mobile CPUs on socket 754; I see, I see... 😁
well done, mate.
Btw disable CnQ and don't use the XP drivers for it, you don't want voltage and frequency scaling when you OC and OV, it's a mobile CPU so you can manually scale it with an utility, just my two cents ofc.

I can add for your reference that my system in Win98SE (Newark at 2933 (11x266) and DDR at 533) in default 1024x768 3dmark01 benchmark scores 20440 on a GF FX 5900 ultra (stock frequencies), which is respectable for my use case.

Edit:
Nvidia Detonator 53.04
Nforce3 Unified Drivers 4.20

Madowax wrote on 2026-04-17, 04:48:

So there were indeed no fun and no reasons in ocing mobile CPUs on socket 754; I see, I see... 😁
well done, mate.

Hmm, for example, my second S754 system, the JetWay K8T8AS, identifies all the Venice cores as Semprons, and, doesn’t enable Cool’n’Quiet on them at all. 😀

My Turion 64 ML-37 arrived so I gave it a quick test. I got it from local laptop spare parts shop, not as cheap as previous ML-34 but not crazy expensive either.

I received 20x20mm copper shims with 0.5mm thickness so I used one on CPU die, with thermal paste on both the die and upper side of the copper shim. Thermaltake MaxOrb EX (full copper) went on top and the mounting screw wasn't tightened all the way, about 1mm remained. Mounting pressure is very good and the issue with bad die contact is gone.

• we have up to 10x CPU multiplier in BIOS
• same as ML-34 at 2.2Ghz Async Latency value 6ns doesn't work so we have to use 7ns. It doesn't reboot but memtest fails within a minute.
• at 2.2Ghz the CPU seems to be running so cool CPUID HWMonitor is unable to take a good reading and reports 18'C. Aida64 shows 2 sensors and one of them reports remperature this low, another one reports more reasonable 25'C when idle. With Super PI 32M running the temperature is 30/29'C on each sensor. Prime95 manages to heat it up to 31/29'C. Not sure if it's a sensor defect or the CPU really runs this cool.
• same as ML-34 Command Rate 1T isn't stable with 2 memory sticks at 2403Mhz CPU, 267Mhz base, 9x multiplier, HT 801 (3x), 400 memory. Super PI 32M fails within a minute.
• at 2.6Ghz and conservative memory timings it does boot into Windows XP but it eventually bluescreens. Quick 5 minute memtest passes. Probably higher voltage than 1.4V is needed, maybe ML-40.

It is clear that getting Turion 64 to run at 2.4Ghz requires Command Rate 2T and more voltage for higher frequency. In my opinion not worth chasing, just get s939 (Toledo) or AM2 (Windsor).

My recommendation would be not to chase Turion 64 ML-37 or faster clocked variants, they are extremely rare, just get ML-34 which are cheap (€10) and run them at 2209Mhz CPU, 260Mhz base, 8.5x multiplier, HT 779 (3x), 400 memory (Command Rate 1T). The CPU is 4% faster than Clawhammer 3400+, consumes less power and has SSE3. My CPU cooler fan runs at 1200 RPM, but you can get a slower one at 800 RPM.

AlexZ wrote on 2026-04-20, 19:18:

[*] at 2.2Ghz the CPU seems to be running so cool CPUID HWMonitor is unable to take a good reading and reports 18'C. Aida64 shows 2 sensors and one of them reports remperature this low, another one reports more reasonable 25'C when idle. With Super PI 32M running the temperature is 30/29'C on each sensor. Prime95 manages to heat it up to 31/29'C. Not sure if it's a sensor defect or the CPU really runs this cool.

The Athlon 64 has two temperature sensors – one is fully internal with a digital interface, and it displays normal readings.
The second is an analogue in-die thermodiod, which is measured by a monitoring chip (Winbond, ITE or similar).
For 90nm CPUs (Venice+), the appropriate coefficients* must be entered in the BIOS.
Some manufacturers released updates with precisely these release notes.
Others ignored them.
Yes, at the upper end of the temperature range, the readings for the thermodiode and the digital sensor converge.
The only drawback is that the fans are controlled by the same monitoring chip that processes the thermodiode.
Therefore, without adjustment, the fan speed is regulated in a very strange way.

* That’s not a table of output-to-input relationships, but the slew rate. Its still need to calculate it from that table.
Many manufacturers couldn’t be bothered to do that.
The fan’s spinning – and that’s that.

It looks like the adjustment is missing in BIOS, it only manifests in low temperatures though.

Turion 64 ML-34@2.2Ghz with with insufficient cooler pressure had 30'C when idle and 50'C in peak load (reached within a few seconds). This proves when temperature rises, it is really reported.
Turion 64 ML-37@2.2Ghz with proper cooler pressure has 24'C when idle and 31'C in peak load. BIOS reports 28'C when idle but it doesn't enter C1 (HLT) like Windows can. ML-37 is also better silicon than ML-34.

I ran Crysis benchmark and got 27.9fps in 1600x1200 with high details except for 4x anti-aliasing quality. Clawhammer 3400+ got about 25fps so we got a small boost here as well. Clawhammer ran with Command Rate 2T, Turion 64 runs 2 double sided DDR400 memory sticks with Command Rate 1T.

Archer57 got 26.7fps in 1920x1080 in s939 Athlon 64 X2 4800+ (2.4Ghz, Toledo) GeForce 8800 GT 1GB. See Re: Socket 939 dual core build. Decisions, decisions.... for results and Re: Socket 939 dual core build. Decisions, decisions.... for system details. GeForce 8800 GT is still underpowered for s939, hence the low score.

It shows that s754 can be surprisingly competitive when paired with an overpowered GPU.

My Socket 754 system comprises of:

* AMD Sempron 64 3400+ processor.
* Asus K8V SE Deluxe (VIA K8T800) Motherboard.
* 1GB DDR400 RAM.
* 256MB PNY Verto Geforce FX 5600 GPU.
* Sound Blaster Audigy 2 ZS sound card for Windows.
* ESS Solo-1 sound card for DOS.
* 80GB + 60GB IDE HDDs from Samsung and Seagate iirc (I have quite a few old IDE drives installed in various systems).
* IDE CD/DVDRW from Samsung.
* Black Jeantech R-2 ATX case (aka GMC/Auzentech R2 Toast).
* 400W FSP400-60GHN ATX Power Supply.
* Creative Inspire T10 Speakers.
* 20" Dell 2007FPB Monitor.
* Dual booting Windows 98SE & XP.

Not used this system for a while as it's currently boxed away amongst a lot of my other old hardware and vintage consoles... It's a case of having too much stuff and not enough space until someone moves out 😂. I originally had it assembled in one of those Commodore CKin ATX cases from back when the Commodore brand briefly dabbled (and failed) in gaming PCs nearly 20 years ago.

I'll have to dig it up at some point and get it going again. I do remember a thread on here about AGP performance issues using the latest BIOS with VIA chipsets and running Windows 98SE, but I don't think my setup was affected afaik (not that I noticed at the time anyway when switching between 98SE & XP).

AlexZ wrote on 2026-04-21, 20:57:

It looks like the adjustment is missing in BIOS, it only manifests in low temperatures though. […]
Show full quote

It looks like the adjustment is missing in BIOS, it only manifests in low temperatures though.

Turion 64 ML-34@2.2Ghz with with insufficient cooler pressure had 30'C when idle and 50'C in peak load (reached within a few seconds). This proves when temperature rises, it is really reported.
Turion 64 ML-37@2.2Ghz with proper cooler pressure has 24'C when idle and 31'C in peak load. BIOS reports 28'C when idle but it doesn't enter C1 (HLT) like Windows can. ML-37 is also better silicon than ML-34.

I ran Crysis benchmark and got 27.9fps in 1600x1200 with high details except for 4x anti-aliasing quality. Clawhammer 3400+ got about 25fps so we got a small boost here as well. Clawhammer ran with Command Rate 2T, Turion 64 runs 2 double sided DDR400 memory sticks with Command Rate 1T.

It shows that s754 can be surprisingly competitive when paired with an overpowered GPU.

In my experience it is better to rise memory frequency over stock than to chase CMD 1T. When the CPU IMC is at such high frequency with more than 1 memory module it might be impossible to reach 1T stable even at lower memory clocks. Relaxed timings at higher memory clock are still very competitive and you can increase the performance gain even more, your choice of mid tier Cpus is the right call imho, you don't need high tier ones for the best results and they are also cheaper.

Yes it may be possible, but requires lot of trial and error so I chose not to do it. The recommended settings should work with minimal effort and little risk.

At 2400Mhz Command Rate 2T seems to be mandatory. To compensate, memory should be OCed as well. Frequency should be chosen after recalculating access time to ns. CAS 2 works at DDR400, but fails very soon at like DDR408. If we switch to CAS 3, frequency should be increased so that access time in ns remains the same.

• DDR400, CAS 2 = 10ns latency
• DDR500, CAS 3 = 12ns latency

Running CAS3 at less than DDR500 makes no sense from latency perspective. Probably 2.8-2.9V will be needed and very good memory. Most people will not do it as they can simply buy s939/AM2.

AlexZ wrote on 2026-04-24, 11:14:

Yes it may be possible, but requires lot of trial and error so I chose not to do it. The recommended settings should work with m […]
Show full quote

Yes it may be possible, but requires lot of trial and error so I chose not to do it. The recommended settings should work with minimal effort and little risk.

At 2400Mhz Command Rate 2T seems to be mandatory. To compensate, memory should be OCed as well. Frequency should be chosen after recalculating access time to ns. CAS 2 works at DDR400, but fails very soon at like DDR408. If we switch to CAS 3, frequency should be increased so that access time in ns remains the same.

• DDR400, CAS 2 = 10ns latency
• DDR500, CAS 3 = 12ns latency

Running CAS3 at less than DDR500 makes no sense from latency perspective. Probably 2.8-2.9V will be needed and very good memory. Most people will not do it as they can simply buy s939/AM2.

I agree but 939 or AM2 are not so good for Win9x, not as good as 754+AGP compatibility wise, for WinXP I prefer CoreDuo, Quad or early Core-I, they are better performer and cheaper.

After installing OCCT 4.4.1, I found out that Command Rate 1T is not 100% stable at 2.2Ghz. OCCT reports error after about 1 hour. Memtest86, prime95, SuperPI 32M worked fine with CR1 even for very long 8+ hour periods. I do have a couple of 1GB DDR400 modules so I will swap them around, but high quality modules like Corsair or OCZ are very rare. nd22 never ran OCCT, so it is possible his 2x 1GB RAM sticks weren't 100% stable at CR1 either. I used 2x Kingston 1GB DDR400 with heatspreader.

This indicates there is some kind of race condition in the CPU that worsens as CPU frequency increases when 2 DDR400 memory sticks are used:

• At 2.4Ghz and CR1, memtest86, prime95 pass but SuperPI 32M fails within a few minutes
• At 2.4Ghz and CR2, memtest86, prime95, SuperPI 32M and OCCT pass
• At 2.2Ghz and CR1, memtest86, prime95 and SuperPI 32M pass, but OCCT fails after 1 hour.

This behaviour was observed in both Turion 64 ML-34 and ML-37 so it isn't just a matter of unlucky silicon.

It is unlikely that one could run at DDR500 or higher with 2x 1GB memory modules in s754. This limits safe usage of CR1 to very expensive memory modules or 1x 1GB DDR400 (win98+winXP dual boot, nForce3/VIA).

When we run at 2.4Ghz, CPU voltage is 1.44V. At 2.2Ghz, CPU voltage is 1.41V. It's nice that it's adjusted up for us.

Gigabyte GA-K8NE is unable to set higher/lower voltage in BIOS.
Asus K8N-E can set lower voltage, but not higher voltage. Higher voltage is ignored and default voltage is used.

Are you running stock and getting errors in Memtest at CR1 with 2 modules?
I tried Superpi 32M and Memtest running overnight and all systems ran 100% stable. I will try OCCT to see if it shows any errors. How does anyone get hold of version 4.4.1?

There are no memtest86 errors at 2.2Ghz CR1. Prime95 and Superpi 32M was stable as well. All tested for very long periods 8+ hours. You need old OCCT version like 4.4.1 to run it in Windows XP and run it overnight.

This problem highlights you need to run at least those 4 stability tests for long periods for CR1. Running only some of them doesn't obviously guarantee stability.

CR1 stability deteriorates with higher CPU clock frequency. It is possible that at 2Ghz it would be fully stable. I will test it and report the results.

OK, where do I find OCCT 4.4.1 for XP?
I let Memtest run overnight for all 4 systems that I tested. Superpi 32m took around half of hour but was stable on all 4 systems also. Maybe you can put up on google drive or something the OCCT version that is compatible with XP so I can test it.

It can be downloaded from http://www.ocbase.com/download/OCCTPT4.4.1.exe. It is the official website for OCCT and they still host old versions. There is also 4.4.2 mentioned in https://www.techpowerup.com/forums/threads/oc … x.142762/page-5

I ran SuperPI 32M repeatedly manually by coming back after about 30 minutes.

Run CPU: OCCT, just click on ON button.

Even with the link provided I still get the latest version and NOT 4.4.1! Internet Archive does not have it!

You will have to google it then. There are several websites still hosting it.

I did a quick test at 2Ghz (default base frequency, default multiplier), CR1, using Auto timings (more conservative than I used for 2.2/2.4Ghz) and OCCT fails within 1 minute or Windows XP doesn't even boot. So this problem is not related to base frequency, CPU frequency, memory frequency or timings. There is some kind of timing issue in the CPU that manifests this strange and may appear stable at certain combination of settings.

To be sure CR1 is stable after running all the tests including OCCT, you should also reduce CPU frequency by 200Mhz and increase base frequency by 1Mhz (for slightly different memory clock) - this will allow to verify you are not in island of stability for CR1 like my 2.2Ghz case. It should be stable for at least 2 CPU frequencies.

The memory instability is due to the IMC that gets OCed with the cpu, so every single addition to the memory integrity path increases instability, best memory high clock or lowest timings chances are with one single memory module. Most of the times you are not going to have a stable system with 2 or more memory slots populated and extreme overclocks.