how to make AMD Turbo not suck?

[SlowSpyder]

cpufreq-utils ftw! Right.

fwiw, my 7700k just doesn't turbo, period. Or at least not that I can identify.

Under the current UEFI rev, the "stock" vcore (if you can call it that) is ~1.212v, and I can get the thing stable at 4 ghz at that voltage, so to heck with turbo. I just lock all the cores at that speed with cpufreq-utils and let it ride. For whatever reason, auto voltage sets vcore to ~1.328v which is just silly.

Make sure your power settings are set to high performance in control panel.

 

[el etro]

It seems firmware 95W TDP and Power Limits are "pushing back" your FX-8310.

 

[Shehriazad]

How to make Turbo not suck:


Step 1: Disable Turbo
Step 2: Manually set the clock you want
Step 3: ?????
Step 4: Profit.

 

[Abwx]

it's hard to be certain about the voltages, they bounce around a lot and not always depending on the frequency-- more intelligent than that-- higher voltage for many heavily loaded cores, lower voltage for fewer loaded cores

Thanks for the pics, thoses values seems to be accurate, at 3.7 base frequency the CPU should be within 95W TDP with 1.275V but that s really the limit when loading is done with Prime 95, that s why i suggested to use Fritzchess bench or eventualy Cinebench, these are more realistic heavy loads, comparitively they would use 75-77W at the seetings where Prime 95 would be at 95W, the turbo behaviour would be different given that the TDP ceiling wouldnt be hit.

 

[DrMrLordX]

run the windows experience index benchmark tool and watch the frequency in HWInfo64.exe. Windows will lock a thread to a core and crank the turbo to the max speed (in my case 4.3ghz) for the duration of the turbo. Cheating. Hrmph!

Which, btw, is 1.413v (1.38v in cpu-z) for me

I might try that if/when I try out Win10 to run some benchmarks on this thing. Right now I'm using Lubuntu.

Make sure your power settings are set to high performance in control panel.

Things are a bit different in Linuxland. The closest I have to that is the performance governor, but at least according to cpufreq-info, the highest clockspeed permitted by the Linux ACPI driver is whatever is set to default.

So when I was running the chip at stock, it had the following available speeds (which correspond to p-states I think):

3.4 ghz
3.1 ghz
2.8 ghz
2.4 ghz
2.0 ghz (1.9 ghz after a BIOS update)

The available states are the same when I'm at 4 ghz, except that it's 4.0 ghz instead of 3.4 ghz at the top-end.

The best realtime clockspeed monitoring I've got thus far is conky-manager, which has a funky non-windowed meter to graphically show current clockspeed. I do not think it could realistically show a turbo clock since the maximum (graph filled) is the current default clockspeed. There is probably something better available, I just haven't got it/installed it yet.

 

[Erenhardt]

I found a way to attach a task to a thread(s) in windows.
You need to have amd Overdrive utility:
http://www.amd.com/en-us/markets/game/downloads/overdrive

in "Performance Control" tab go to "AMD smart profiles"
Add your application and create a profile. You can choose "CPU core affinity" and increase the overclock on that core.

I played today a couple hours with my fx6300 and noticed a couple of things:
1. CPU overclocked to 4.2 and with disabled Turbo have higher performance than with 4.2ghz turbo enabled. (1.10 vs 1.00 CB11.5)

2. I was able to enable higher turbo (4.8GHz) on already overclocked processor without having stability issues encountered while overclocking to the turbo speed.

3. Doing #2 I increased my single thread score by couple %. Multithreaded stayed the same. Its good way to decrease the lack in singlethreaded performance when you motherboard cant push whole CPU to those speeds.

4. setting affinity in CB to single core made system very unresponsive.

...5. My brothers fx6300 overclocks more (4.4 base and 5.0 turbo vs mine 4.2 base and 4.8 turbo)

 

[Headfoot]

but....that's not what I want...

Why?

It's completely pointless to want it to turbo on its own to full frequency instead of just setting it at max frequency under all loads for a desktop system. The difference in power consumption is so small in desktop terms that it really doesn't matter. Might've mattered on a battery

Or go ahead and waste hours upon hours trying to fix this issue that's easily solved with the above

 

[soccerballtux]

Thanks for the pics, thoses values seems to be accurate, at 3.7 base frequency the CPU should be within 95W TDP with 1.275V but that s really the limit when loading is done with Prime 95, that s why i suggested to use Fritzchess bench or eventualy Cinebench, these are more realistic heavy loads, comparitively they would use 75-77W at the seetings where Prime 95 would be at 95W, the turbo behaviour would be different given that the TDP ceiling wouldnt be hit.

you know it bounces around so much it's hard to say if that 1.275 was accurate. Seriously, it bounces a lot.

Check back in a few minutes I'll have a CB SS, but honestly I don't know how I'm going to catch cpu-z while it's rendering, that'll throw it off

 

[soccerballtux]

I might try that if/when I try out Win10 to run some benchmarks on this thing. Right now I'm using Lubuntu.



Things are a bit different in Linuxland. The closest I have to that is the performance governor, but at least according to cpufreq-info, the highest clockspeed permitted by the Linux ACPI driver is whatever is set to default.

So when I was running the chip at stock, it had the following available speeds (which correspond to p-states I think):

3.4 ghz
3.1 ghz
2.8 ghz
2.4 ghz
2.0 ghz (1.9 ghz after a BIOS update)

The available states are the same when I'm at 4 ghz, except that it's 4.0 ghz instead of 3.4 ghz at the top-end.

The best realtime clockspeed monitoring I've got thus far is conky-manager, which has a funky non-windowed meter to graphically show current clockspeed. I do not think it could realistically show a turbo clock since the maximum (graph filled) is the current default clockspeed. There is probably something better available, I just haven't got it/installed it yet.

that was one of the reasons I stopped dealing with linux. Fan control and software overclocking. Although, there is an updated phenommsrtwkr for your FX, google around for it, it was released 2013. It's CLI only but works

 

[soccerballtux]

I found a way to attach a task to a thread(s) in windows.
You need to have amd Overdrive utility:
http://www.amd.com/en-us/markets/game/downloads/overdrive

in "Performance Control" tab go to "AMD smart profiles"
Add your application and create a profile. You can choose "CPU core affinity" and increase the overclock on that core.

I played today a couple hours with my fx6300 and noticed a couple of things:
1. CPU overclocked to 4.2 and with disabled Turbo have higher performance than with 4.2ghz turbo enabled. (1.10 vs 1.00 CB11.5)

2. I was able to enable higher turbo (4.8GHz) on already overclocked processor without having stability issues encountered while overclocking to the turbo speed.

3. Doing #2 I increased my single thread score by couple %. Multithreaded stayed the same. Its good way to decrease the lack in singlethreaded performance when you motherboard cant push whole CPU to those speeds.

4. setting affinity in CB to single core made system very unresponsive.

...5. My brothers fx6300 overclocks more (4.4 base and 5.0 turbo vs mine 4.2 base and 4.8 turbo)

that's interesting, thanks. So, todo, write a script that makes a custom profile for every single-threaded application I run, to set affinity to the turbo'd core.

4. maybe windows set an affinity for the scheduling process to be on core 1?

5. try turning off 2 of your cores and overclocking again.

 

[soccerballtux]

Why?

It's completely pointless to want it to turbo on its own to full frequency instead of just setting it at max frequency under all loads for a desktop system. The difference in power consumption is so small in desktop terms that it really doesn't matter. Might've mattered on a battery

Or go ahead and waste hours upon hours trying to fix this issue that's easily solved with the above

I understand what you're saying because I said it myself many times, but honestly I like the lower TDP/heat/power. I had it overclocked on my Ph2-965, would downclock it to stock, it just didn't feel as fast as this does.
maybe that's the fresh format I've got

 

[soccerballtux]

5.54pts in Cinebench 11.5.
Spent most of its time at 3.4ghz

20k in CB10-- spent most of its time at 3.7ghz.

what I really need is some logging tool to check the voltage/freq every 200ms

 

[soccerballtux]

edit: so uh, somehow overdrive was interfering with settings, or I needed overdrive specifically in order to enable Turbo Core manually better or something weird, but now they're all turboing to 4.3ghz no problem, and somehow at 4.3ghz my superpi time is 22s, when it used to be 17s on my Ph2-965 @ 4Ghz and 2.6Ghz CPU-NB.

hmph.

no, that's not what appears to have gone on either.

here's stock without overdrive:

here's with the turbo setting enabled in overdrive

this way works pretty nicely, turbos all the time, very snappy

 

[nenforcer]

I thought about going the AMD FX-8300 / 8310 / 8320E / 8370E route but those Max Turbo speeds are misleading. You normally only achieve the max turbo (4.2GHz / 4.3GHz) with less than half the cores 4 / modules (2) loaded and like half the turbo (3.7GHz) with more than 2 modules (4) cores but less than a fully loaded (4 modules) / 8 cores which runs at exactly the base clock to maintain the lower 95W TDP.

I think I've decided to just get the 8350 / 8370 and maintain the higher base clock at a 125W TDP because I want all 8 cores clocked at that higher speed.

The AMD FX 9370 / 9590 are really just diminishing returns at a 220W TDP, however.

 

[DrMrLordX]

that was one of the reasons I stopped dealing with linux. Fan control and software overclocking. Although, there is an updated phenommsrtwkr for your FX, google around for it, it was released 2013. It's CLI only but works

msrtweaker for teh lunix? huh, interesting. I know there's turionpowercontrol and it supports Piledriver, but not Kaveri (I know because I've tried it). There's a version I see mentioned on some Italian forums that does support Kaveri, but I can't find a link to it. If I did actually have an FX, I could just use turionpowercontrol but oh well.

Regardless I have defeated the evil Kaveri throttle monster so I'm satisfied with that.

On a side note regarding p-state throttling and turbo: one revealing bit of information that cpufreq-utils provides via the 'cpufreq-info' command is that the maximum transition latency is 4 microseconds. I'm not sure if a core goes idle or continues operating normally while it experiences clockspeed transition latency, but still, in computing terms, 4 microseconds is a loooong time.

 

[Erenhardt]

It seems that windows is not always boosting the cores that have the most load on them.

Running a single threaded benchmark the core load jumps between threads. So does turbo. But it looked like the loaded thread is not always the one that runs on turbo.

I was going through my applications wondering what would use single thread to see exactly if the turbo is working as it should - boosting the most loaded thread.

And I found that Torchlight 2 loads single thread (#0) and stays there. Turning turbo on and off seems to impact performance a bit. 4.2Ghz base, 4.6Ghz turbo seems to be adding around 5-10% more fps (game entirely CPU bound, with 40% GPU utilization).

But I noticed on my OSD that the boosted core is not always the one under load.

 

[soccerballtux]

I thought about going the AMD FX-8300 / 8310 / 8320E / 8370E route but those Max Turbo speeds are misleading. You normally only achieve the max turbo (4.2GHz / 4.3GHz) with less than half the cores 4 / modules (2) loaded and like half the turbo (3.7GHz) with more than 2 modules (4) cores but less than a fully loaded (4 modules) / 8 cores which runs at exactly the base clock to maintain the lower 95W TDP.

I think I've decided to just get the 8350 / 8370 and maintain the higher base clock at a 125W TDP because I want all 8 cores clocked at that higher speed.

The AMD FX 9370 / 9590 are really just diminishing returns at a 220W TDP, however.

I'm actually having issues now where only all 8 cores are Turboing, this is restricting me to P-state Pb1 at 3.7ghz.

This might be a mobo bug.
I need to find a hack around it.

 

[soccerballtux]

msrtweaker for teh lunix? huh, interesting. I know there's turionpowercontrol and it supports Piledriver, but not Kaveri (I know because I've tried it). There's a version I see mentioned on some Italian forums that does support Kaveri, but I can't find a link to it. If I did actually have an FX, I could just use turionpowercontrol but oh well.

Regardless I have defeated the evil Kaveri throttle monster so I'm satisfied with that.

On a side note regarding p-state throttling and turbo: one revealing bit of information that cpufreq-utils provides via the 'cpufreq-info' command is that the maximum transition latency is 4 microseconds. I'm not sure if a core goes idle or continues operating normally while it experiences clockspeed transition latency, but still, in computing terms, 4 microseconds is a loooong time.

CLI as in, command prompt.
AMDMsrTweaker

 

[soccerballtux]

It seems that windows is not always boosting the cores that have the most load on them.

Running a single threaded benchmark the core load jumps between threads. So does turbo. But it looked like the loaded thread is not always the one that runs on turbo.

I was going through my applications wondering what would use single thread to see exactly if the turbo is working as it should - boosting the most loaded thread.

And I found that Torchlight 2 loads single thread (#0) and stays there. Turning turbo on and off seems to impact performance a bit. 4.2Ghz base, 4.6Ghz turbo seems to be adding around 5-10% more fps (game entirely CPU bound, with 40% GPU utilization).

But I noticed on my OSD that the boosted core is not always the one under load.

yes, that's frustrating, poor implementation.

I think what would be a lot simpler is if they just turbo'd all the cores and turbo'd down to stay within the TDP budget. So, 95W, happens to work out to about 3 cores at full turbo; 6 at half turbo, 8 at no turbo