Godavari throttles to 1.6 GHz after 20 seconds

[JDG1980]

One of the issues hurting performance of Kaveri APUs was the aggressive throttling. Unfortunately, it looks like this hasn't been fixed in Godavari, and may have gotten worse; according to tests run by CPU World, running the A10-7870K at full blast for over 20 seconds results in the CPU throttling back to a mediocre 1.6 GHz. That doesn't seem to be due to TDP violations, as it happens even in a single-threaded SuperPi test (presumably with low or no GPU load). How could AMD have dropped the ball this badly?

 

[FlanK3r]

the operating voltage is crazy high, nonsense. I think, BIOS need new agesa.

 

[CHADBOGA]

One of the issues hurting performance of Kaveri APUs was the aggressive throttling. Unfortunately, it looks like this hasn't been fixed in Godavari, and may have gotten worse; according to tests run by CPU World, running the A10-7870K at full blast for over 20 seconds results in the CPU throttling back to a mediocre 1.6 GHz. That doesn't seem to be due to TDP violations, as it happens even in a single-threaded SuperPi test (presumably with low or no GPU load). How could AMD have dropped the ball this badly?

Surely a quick fix will be imminent or else they will have to do extensive benching only with GeekBench. :biggrin:

 

[LTC8K6]

http://cdn.overclock.net/6/60/900x900px-LL-60e22ea6_A10_7870K_06Cpuz.png

Same chip, same version of CPU-Z

4.7Ghz at 1.45V

 

[LTC8K6]

Overclock.net had the chip at stock speeds with a much lower voltage than the 1.48v the board defaulted to.

So it seems like a default voltage problem.

 

[monstercameron]

That sucks, had my fill of kaveri and its 3ghz throttling behavior.

 

[Shehriazad]

Some boards default to ridiculous voltages.

Kaveri needs to be tweaked manually for sure, it doesn't even seem to be a heat problem...because throttling could occur solely when the iGPU is drawing too much juice (or maybe it's heat related because the iGPU is right next to the CPU which causes instant heat that isn't properly registered by sensors? Not sure, not an engineer).

Maybe it's an issue with Socket FM2+.


That said...people ran their 7850Ks as low as 1.2ish Volts for 4 Ghz(The LOWEST I ever saw recorded with proof was 1.188). 1.25 to 1.3 should be EASY for perma 4.1

And since that perma 4.1 is gonna be volted lower than auto volt the APU would even have a longer lifespan (in theory)
The iGPU also gets ridiculous Voltages at times. Some boards smash it up to 1.4...but again 1.2 is nice enough for a decent OC + 2400 Mhz ram.


So I'd guess 7870K would on average run fine on every PC with 1.275 + 1.2 Vcore and there would be no more throttling.


I am 100% unsure why a majority of boards overvolts Kaveri/Godavari SO harshly when it is absolutely not necessary. Why are they auto volting to 1.4 to 1.48 if they are stable SO much lower than that. Is AMD to blame or is it the board makers? I'm not even sure at this point....but it feels like self-sabotage.

 

[LTC8K6]

First of all, the 7870K runs happily off lower voltages than the default 1.487V for 3.9GHz and I was able to run 3.9GHz from below 1.2V.

Rather than sticking with multiples of 100, which CPUZ always used to reduce by a few hertz, just to make you miss a GHz target, I planned my 4.5GHz-4.9Ghz series using clearance multiples (43x105=4515, 44x105=4620, 44x107=4708, 45x107=4815, 45x109=4905). The first three worked fine: 4.515GHz=1.43125V, 4.620GHz=1.45V and 4.708GHz=1.475V; but 4.815GHz crashed with voltage raised to 1.5375V, even with the LLC settings set to extreme, so my 'Hello Boys, I'm Back' poster will have to stay unposted and I will have to leave off testing this chip any further until Asus update the BIOS and/or someone lets us know what the upper limit for CPU voltage is.

http://www.overclock.net/t/1555615/amd-kaveri-refresh-a10-7870k-apu-arrives

 

[DrMrLordX]

The default voltages are insane. The pstates are insane. Turn down the voltage, bam, problem solved.

Clueless end-users who do not understand these things may have problems, but reviewers darn well ought to know better. My guess is it's a board problem. It comes down to the UEFI what default voltages are detected (I've seen that change between UEFI revs on my A88x-Pro, for example).

All signs point to these chips (at least the new-stepping ones, which you get in the 7870k and 7670k) hitting 4.7 ghz pretty regularly. They're a step up from older Kaveri chips.

 

[LTC8K6]

Two different reviewers with two different boards had the same default, 1.487v for the 7870K.

 

[myocardia]

Two different reviewers with two different boards had the same default, 1.487v for the 7870K.

I'm not surprised. Make the minimum VID the same as the maximum, and you'll have almost zero die that can't 'make the cut'. It makes sense, when you are losing money every quarter, to lose as little as possible.

 

[freeskier93]

Stock voltages are certainly... aggressive. My older A8-5600k was running at 1.4 something volts for stock 3.6-3.9 Ghz clock. I get a stable 4.5 Ghz at 1.31 volts.

 

[frozentundra123456]

I'm not surprised. Make the minimum VID the same as the maximum, and you'll have almost zero die that can't 'make the cut'. It makes sense, when you are losing money every quarter, to lose as little as possible.

If they are in fact doing this to make yields better, they are shooting themselves in the foot basically. Yes, they save money on yields, but the vast majority of users are not going to be aware of this problem, and even if they were, I am not sure OEM boards would allow you to tweak the voltage. So they are selling a product that performs less well than it should, losing user satisfaction and possible future purchases, to save a few dollars on the cost of the chip.

Personally though, I think this just sloppy design, not done intentionally to make more chips pass the cut. Question is, who is responsible, AMD or the motherboard makers? And even if it is the mb makers, why isnt AMD addressing this problem?

 

[JDG1980]

The default voltages are insane. The pstates are insane. Turn down the voltage, bam, problem solved.

They may be insane, but they are the defaults. Many users aren't going to touch them (especially in OEM systems, if there are any). You shouldn't have to tinker around in the BIOS to get the CPU to work the way it's supposed to. Out-of-the-box settings should work adequately, and be devoid of blatant flaws like the throttling observed here.

 

[VirtualLarry]

If they are in fact doing this to make yields better, they are shooting themselves in the foot basically. Yes, they save money on yields, but the vast majority of users are not going to be aware of this problem, and even if they were, I am not sure OEM boards would allow you to tweak the voltage. So they are selling a product that performs less well than it should, losing user satisfaction and possible future purchases, to save a few dollars on the cost of the chip.

Personally though, I think this just sloppy design, not done intentionally to make more chips pass the cut. Question is, who is responsible, AMD or the motherboard makers? And even if it is the mb makers, why isnt AMD addressing this problem?

I recall reading a "decision tree" for VID on AMD CPUs, and basically, the logic was faulty, and resulted in VID == Max VID every time.

So if someone at AMD came up with this faulty logic tree for VID calculations, and passed it on to mobo makers, and they implemented it verbatim, then ... that's what you get.

 

[ShintaiDK]

You would think it was an issue that would be addressed after the blunder with the first generation of Kaveri chips and their relatively mild throttle compared way below baseclock to the 3Ghz mark. But 1.6Ghz....

 

[mrmt]

I'm not surprised. Make the minimum VID the same as the maximum, and you'll have almost zero die that can't 'make the cut'. It makes sense, when you are losing money every quarter, to lose as little as possible.

So basically godavari is trash silicon that couldn't make it as Kaveri?

 

[DrMrLordX]

They may be insane, but they are the defaults. Many users aren't going to touch them (especially in OEM systems, if there are any). You shouldn't have to tinker around in the BIOS to get the CPU to work the way it's supposed to. Out-of-the-box settings should work adequately, and be devoid of blatant flaws like the throttling observed here.

OEM systems are predominantly getting SKUs from older steppings. The only OEM-designated CPU from the new stepping is the 8850B:

http://wccftech.com/amd-a10-7870k-black-edition-godavari-flagship-retail-150/

Regardless, OEMs aren't going to stick these chips in UEFI-updated retail boards and ship OEM machines with stock vcore settings of 1.4v and higher. They're going to go back, check their own work, notice the goofy-ass voltage tables, and patch their in-house UEFI to bring chip voltage down to at least 7850k levels to bring power consumption within spec, possibly after making some angry phone calls to AMD about voltage issues.

OEMs like to cut corners on cooling. They will notice the problem on the 8850B and fix it before units ship. Otherwise their units just won't work as designed. There will be a ton of thermal throttling.

I recall reading a "decision tree" for VID on AMD CPUs, and basically, the logic was faulty, and resulted in VID == Max VID every time.

So if someone at AMD came up with this faulty logic tree for VID calculations, and passed it on to mobo makers, and they implemented it verbatim, then ... that's what you get.

. . . well great. That sounds about right, which is sad but what can you do? Let the mobo manufacturers roll out some UEFI patches and get things working properly. In the meantime, the DIY crowd (which is the only crowd that will have this problem, really) can just turn down the darn voltages by hand. Kaveri already has a rep for auto overvolts for whatever stupid reason, so it's more of the same there.

So basically godavari is trash silicon that couldn't make it as Kaveri?

Uh, no. Kaveri refresh has two different steppings: the older KA-V1 that will probably behave the same as later-gen KA-V1 chips such as the 860k and 7650k, and the newer stepping GV-A1 chips that will reach the same clocks at lower voltages compared to KA-V1. GV-A1 chips (8850B, 7670k, 7870k) are also more likely to break 4.5 ghz than KA-V1, though they will probably be no more likely to eclipse 4.7 ghz on air/water than the older stepping.

I predicted this nonsense when Himo5 first posted the p-states for his 7870k on OCN.

By the by, a throttle to 1.6 GHz under extreme overvolt conditions (which is what we see from some of these defaults) has nothing to do with the CPU throttling to 3 ghz under iGPU load observed in Kaveri since the beginning (unknown if Kaveri refresh still does that). I got throttling similar to that when my temps went out of control after my first unsuccessful attempt to relid my CPU after delidding. If you take an unaltered Kaveri chip and push too much voltage, you can overheat the thing and produce the same effect.

 

[JDG1980]

I recall reading a "decision tree" for VID on AMD CPUs, and basically, the logic was faulty, and resulted in VID == Max VID every time.

So if someone at AMD came up with this faulty logic tree for VID calculations, and passed it on to mobo makers, and they implemented it verbatim, then ... that's what you get.

How does something like that get past QA?

Regardless, OEMs aren't going to stick these chips in UEFI-updated retail boards and ship OEM machines with stock vcore settings of 1.4v and higher. They're going to go back, check their own work, notice the goofy-ass voltage tables, and patch their in-house UEFI to bring chip voltage down to at least 7850k levels to bring power consumption within spec, possibly after making some angry phone calls to AMD about voltage issues.

OEMs like to cut corners on cooling. They will notice the problem on the 8850B and fix it before units ship. Otherwise their units just won't work as designed. There will be a ton of thermal throttling.

I think you give OEMs too much credit. Most won't care about thermal throttling. The chips run, they don't crash, for run-of-the-mill, penny-pinching OEMs this will be good enough.

 

[escrow4]

If an everyday user buys this, they won't touch the volts (they wouldn't know what volts is likely) but they will likely notice the slowdowns. Poor form AMD. Then again, an everyday user would probably buy Intel anyway. Perception and all that. Not surprising. AMDs FX CPUs are a joke, and their APUs are a tiny niche. How could you release a product like this?

 

[Shehriazad]

Godavari seems to overclock a little better...and the Athlon 860K also ran with lower voltage for me...so I'm waiting for the 870K now.

I'm hoping to push it closer to 5Ghz than the 860K. I'm hoping that it can max out at like 4.8 with some cherry picks running 4.9.


But seeing how Godavaris APUs seem to hit 4.7 "often" now (Something that wasn't quite guaranteed with Kaveri) I hope that this is at least something. I mean the "serial naming" did indicate a new stepping (JC)...so yea.

 

[naukkis]

Throttle all the way to 1.6GHz means that MB voltage regulation overheat and protection kicks in.

 

[Shehriazad]

Throttle all the way to 1.6GHz means that MB voltage regulation overheat and protection kicks in.

But who exactly is at fault for almost all of the mainboards overvolting so freakin' massively?

I mean even the 860K gets standard volted to like 1.41 on my board. My 860K also runs 4.1 Ghz @ 1.2V.

My old 7850K also managed 4.0 @ 1.2.


Even if you want to make sure that every chip can run and thus give it some extra juice...it seems like even 1.3 would have been just fine.

When I first got that Asrock mainboard (sig) I even asked support what's up with that...and they said it was a security measure to ensure stability...but that just seems so ridiculously over the top. AMDs own overdrive tool even shows the standard voltages as deeeep red zone..and that makes sense.

 

[PG]

I have an 860K and cheap Asus A88XM-A. After a recent bios update I noticed that under load, at stock speed, the voltage was hitting 1.512. It was insane.
Manually I can overclock to 4.3 Ghz and undervolt to just under 1.4V at load and it runs fine.
I have no idea why most boards now overvolt like crazy but it's a real problem and one I hope AMD and the board manufacturers address soon.
Bottom line: watch the default voltage on any FM2+ board with a recent bios update and adjust down as necessary, even if not overclocking

 

[DrMrLordX]

How does something like that get past QA?

It seems like they are doing it deliberately, though it's impossible to say exactly why. We may be seeing an undesirable side-effect of AMD cutting jobs and trying to run everything on a shoestring budget. Neither Kaveri nor Kaveri refresh require anywhere near that much voltage for stock operation. Every Kaveri chip ever released can do stock speeds with between 1.2-1.25v, and should be able to achieve the full range of turbo operation with 1.35v or less.

The only thing I can think of is that there are some motherboards out there that have really loose voltage regulation, particularly when the CPU enters or leaves full load state. Kaveri tends to be picky about voltage regulation. If there's too much droop, the chips can get unstable, even if the droop is only momentary. The heavy overvolting may be a preventative measure for whatever percentage of boards out there are shipping with marginal VRMs that have abnormally poor vdroop going into a load.

That being said, Asus is the brand that seems to have the best VRMs among all the FM2+ board OEMs right now, and they're pushing insane voltage too.

It is a simple matter of fact that all the OEMs pushed high voltages for Kaveri from the start, and for whatever reason, they're starting to push more voltage now, especially on newer chips.

I think you give OEMs too much credit. Most won't care about thermal throttling. The chips run, they don't crash, for run-of-the-mill, penny-pinching OEMs this will be good enough.

That might fly for some mobile products where clockspeeds are expected to jump around. Surely you're thinking of stuff like the Yoga 3? For desktops, generally speaking, constant throttling is abnormal and will not pass muster. Heck even the cheapo Jaguar AiO desktops you can get from big box stores don't throttle. They run kinda warm, but you can set the performance profile in Windows and lock them into max frequency and they'll never drop for anything. Even cheap-arsed OEMs won't ship a desktop unit that misses that mark. They will ship one that's noisy and that runs at too-high a temperature, but . . .

But who exactly is at fault for almost all of the mainboards overvolting so freakin' massively?

I mean even the 860K gets standard volted to like 1.41 on my board. My 860K also runs 4.1 Ghz @ 1.2V.

My old 7850K also managed 4.0 @ 1.2.


Even if you want to make sure that every chip can run and thus give it some extra juice...it seems like even 1.3 would have been just fine.

When I first got that Asrock mainboard (sig) I even asked support what's up with that...and they said it was a security measure to ensure stability...but that just seems so ridiculously over the top. AMDs own overdrive tool even shows the standard voltages as deeeep red zone..and that makes sense.

Something isn't right . . . and the idea that these chips need 1.4-1.5v "as a precaution" is a complete joke. The board OEMs are probably saying stuff like this to save face after they knowingly implemented a flawed voltage table from AMD. Unless they are criminally insane, they'll fix it quietly in the weeks to come.

I have an 860K and cheap Asus A88XM-A. After a recent bios update I noticed that under load, at stock speed, the voltage was hitting 1.512. It was insane.
Manually I can overclock to 4.3 Ghz and undervolt to just under 1.4V at load and it runs fine.
I have no idea why most boards now overvolt like crazy but it's a real problem and one I hope AMD and the board manufacturers address soon.
Bottom line: watch the default voltage on any FM2+ board with a recent bios update and adjust down as necessary, even if not overclocking

If what you are saying is correct, then even people buying chips with the old stepping will need to be careful. Of course, really, everyone buying these chips should be turning down voltage anyway. They overvolts have been happening for awhile. They're just getting worse.