PSU tier list 2.0

[Karadjgne]

If you look at it, it's basically still nothing new. 7 or so years ago when Seasonic released the 12ii models, they were toted as mythical quality, according to the time. With some of today's units in comparison, they barely break mediocre for output, while still good quality. With ATX standards getting stricter with each revision, any 'good' unit from 10 years ago would be well out of compliance now. I'd hate to see how an old BFG would stack up, even though they were top(ish) line back then.

 

[goldstone77]

OW wolf would never ever go back to old units and change scores. Don't count on it in a million years. For one thing, new PSUs are scored stricter because the standards for what is "good" is no longer the same. So when you see a 9.0 score that's a 9.0 score for whatever year the review it was, and you always have to keep that in mind. Also, hope and pray for the best of his father, he has cancer so he's been super busy with helping out his father.

I remember reading something to the effect of the new high end units are so good they would all be 10's so he had to change some things to make it harder.

"Mythic quality" is what he says they call new high quality PSUs on his website.

I saw that. 😀

They just keep getting better all the time, the high quality ones anyway.

I did notice how careful he was being about what he said in the video, especially when it came to Corsair PSU questions.

Yeah, and if you notice he gives you a hint about how to find a good PSU "10 year warranty" from brand name manufactures. He mentions Corsair units that come with 10 year warranty have quality components expected to last 24/7 operation for 10 years.

 

[Sakkura]

If you look at it, it's basically still nothing new. 7 or so years ago when Seasonic released the 12ii models, they were toted as mythical quality, according to the time. With some of today's units in comparison, they barely break mediocre for output, while still good quality. With ATX standards getting stricter with each revision, any 'good' unit from 10 years ago would be well out of compliance now. I'd hate to see how an old BFG would stack up, even though they were top(ish) line back then.

The ATX standard hasn't really changed much in recent years. Most of the progress is just people wanting higher and higher quality. We're still pretty much running ATX v2.31 (of 2008) where it counts.

Even Intel sort of acknowledged that by having a few minor revisions after 2.31 that they didn't even put a separate number on. They did eventually go back to numbering with 2.4, but that's still a minor revision.

 

[Karadjgne]

I stuck some old CX430 in some office pc's that even after 5 years were still going strong, not a single bunk unit out of 10 and 0 complaints from the owner. Stuck a CX500 in a 'gaming' rig (this was back in the days of Win98se before WindowsMe even surfaced) that didn't last 6 months before it was toast. Warranty age is nice, but really only applicable to the right unit for the right application. I'd not stick a CX850 in a high-end gaming rig, but have no issues using one in a 10hdd small office Franken-server

 

[jankerson]

OW wolf would never ever go back to old units and change scores. Don't count on it in a million years. For one thing, new PSUs are scored stricter because the standards for what is "good" is no longer the same. So when you see a 9.0 score that's a 9.0 score for whatever year the review it was, and you always have to keep that in mind. Also, hope and pray for the best of his father, he has cancer so he's been super busy with helping out his father.

I remember reading something to the effect of the new high end units are so good they would all be 10's so he had to change some things to make it harder.

"Mythic quality" is what he says they call new high quality PSUs on his website.

I saw that. 😀

They just keep getting better all the time, the high quality ones anyway.

I did notice how careful he was being about what he said in the video, especially when it came to Corsair PSU questions.

Yeah, and if you notice he gives you a hint about how to find a good PSU "10 year warranty" from brand name manufactures. He mentions Corsair units that come with 10 year warranty have quality components expected to last 24/7 operation for 10 years.

And he also said the RMX series was the best in terms longevity in the Corsair line, that says a lot, 10 year warranty on those.

Also noted was the use of adaptors as in don't use them EVER.

And some of the other issues that people create for themselves that we see here on the forums everyday.

 

[jankerson]

If you look at it, it's basically still nothing new. 7 or so years ago when Seasonic released the 12ii models, they were toted as mythical quality, according to the time. With some of today's units in comparison, they barely break mediocre for output, while still good quality. With ATX standards getting stricter with each revision, any 'good' unit from 10 years ago would be well out of compliance now. I'd hate to see how an old BFG would stack up, even though they were top(ish) line back then.

Yeah there have been a lot of changes since then.

Taking a 7+ YO design and comparing it to a current one isn't very realistic.

That said I wouldn't use a 7+ YO design in a current machine with all of the latest hardware.

 

[Karadjgne]

Oh I would. Gpus haven't really changed in design, just architecture, but the rest of the components are still the same caps, same vram, same VRM's etc as was used 7 years ago. I'd have no qualms about putting my 5yr old Seasonic M12-II 520w semi up against a gtx1060 build. Hardware today is still built to withstand much worse outputs than what that psu supplies. I'd take that M12-II over a brand new TR2 any day of the week.

 

[jankerson]

Oh I would. Gpus haven't really changed in design, just architecture, but the rest of the components are still the same caps, same vram, same VRM's etc as was used 7 years ago. I'd have no qualms about putting my 5yr old Seasonic M12-II 520w semi up against a gtx1060 build. Hardware today is still built to withstand much worse outputs than what that psu supplies. I'd take that M12-II over a brand new TR2 any day of the week.

Well yeah, but you know what I was talking about.

The M12ii isn't a junk PSU, even by today's standards, and the 750W model and up are Haswell Certified.

The 620W and below aren't, and that includes the EVO's. (DC-DC)

 

[Karadjgne]

The only part thats incompatible with Haswells is the deep sleep states beyond c-3. Turn that off in either windows power or bios settings and there's no reason you can't use a group regulated design for any cpu. And they work just fine for Ryzen as is, so it's only really a small portion of new cpus that are affected worldwide, lga1155 is still very popular in Asian markets.

 

[Darkbreeze]

Yeah, I don't know about that. Although the Haswell implementation is when it became a well known and popular factor, Intel has been using the C6/C7 states since the 1st gen core-i series i7-900 and unless there is something wildly different in the implementation of those states on earlier architectures, they should experience the exact same problems with the group regulated models that Haswell and newer core-i models do.

Probably there are, or were, issues with these systems related to that and that relationship might just have not been known. I've been trying to find something that shows the implementation is different or that definitely proves there is no such issue on earlier architectures, but I cannot find anything that does that. I see no reason why the problem wouldn't have existed on these earlier models too and my guess is that it did, but nobody knew that was the reason for these problems then. Perhaps not, but I can't find any evidence to the contrary.

I even reached out to Seasonic on this and they could not give me an answer one way or the other except to say that units which do not say they are Haswell compatible should have this problem on any platform using C6/C7 low power settings.

Section 7.3 shows that this architecture uses those states, and all the succeeding architectures show it is used as well.

https://www.intel.com/content/www/us/en/processors/core/core-i7-900-ee-and-desktop-processor-series-datasheet-vol-1.html

 

[Karadjgne]

Group regulated designs in low power mode only go down so far. Below that limit, there's not enough juice to enable the psu to break out of the deeper sleep modes set by the cpu. Pre-4th gen low power modes like C-6/7 didn't broach that low power limit, so when something woke the cpu, the psu went back to full function. Haswell and newer Intel have very low power needs in C-6/7, actually less than what a group regulated psu can function at, so when a Haswell hits C-6, the psu simply won't come out of sleep. DC-DC designs limits are considerably lower than group regulated, so don't suffer the same limitations.

The sleep states are the same, it's not a set voltage value it's a functionality state. The actual voltages used by the cpus are different.
C0 Operating State CPU fully turned on
C1 Halt Stops CPU main internal clocks via software; bus interface unit and APIC are kept running at full speed.
C1E Enhanced Halt Stops CPU main internal clocks via software and reduces CPU voltage; bus interface unit and APIC are kept running at full speed.2 Stop Grant Stops CPU main internal clocks via hardware; bus interface unit and APIC are kept running at full speed
C3 Sleep Stops all CPU internal clocks
C4 Deeper sleep Reduces CPU voltage
C5 Enhanced Deeper Sleep Reduces CPU voltage even more and turns off the memory cache
C6 Deep Power Down Reduces the CPU internal voltage to any value, including 0 V

So it all depends on the cpu voltages, Haswell and newer dropping down below the threshold a group regulated design can supply at C-4 and below. Haswell cpus go down to C-10 I believe, each state below C-5 just being progressive CPU voltage lowering, the above C-6 being Haswell C-10 equivalent.

In my Asus bios, it's set as S1 or S3, S1 being limited to C0-C3, S3 is full function C0-C6. Default is S3, but that's OK as it's a 3rd gen cpu so doesn't go below the threshold of my M12-II at C4-C6, if it was a Haswell cpu I'd need to stick with S1, which allows sleep mode, but not deep sleep where cpu voltages are actually lowered.

 

[jankerson]

My 1st gen i7 870 is set to S3 by default and it works fine, but the machine has a Seasonic G-550W in it and it comes right back out of it with no issues.

The machine is like it's turned off, no power, lights, nothing.

 

[Darkbreeze]

How is C6/C7 implemented any differently on 4th-8th gen than on 3rd gen, or 1st gen for that matter? I've asked around to all of the power supply manufacturer/brand tech support and through many other sources, and can't seem to find ANY definitive answer that there is any difference/change in the specification or implementation of the specification. I see the same S/sleep states, down to S0 and low power C states, down to C6/C7 on EVERY core-i CPU from 1st to 8th gen. There is no differentiation in ANY of the Intel documentation regarding these states or their implementation. I've read through every single Core-i datasheet, looked at all the diagrams of core and package power state implementation and design, and I don't see ANY difference in them at all.

So again, I fail to see how pre-4th gen Core-i processors handle those states any differently than 4th gen and newer models do, which to ME says they should be affected exactly the same as any other processor with this issue regarding the low power conflict on group regulated power supplies.

 

[Karadjgne]

Because they don't handle it differently. They handle it exactly the same. With different values. Just to plug on numbers, if a 3rd gen has a working range where absolute low cpu voltage is 1v, 4th gen absolute low voltage would be something like 0.3v. A group regulated psu might have a 0.8v threshold, a dc-dc 0.2v threshold. They work exactly the same for C states, S states and P states, the difference being the cpu voltage values. At C0-C3 you are still looking at normal cpu voltages, but at C4 the cpu lowers its requirements to maintain clocks as the clocks are not on. That trickle voltage on a Haswell cpu is just lower than Ivy, dropping down to 0.7v at C4. That's plenty for a dc-dc, but going below the group regulated threshold.
The actual process is the same, but being analog, the group regulated psu just can't remain in an operating ready state if the cpu voltage drops too low. There's simply not enough voltage in the cpu to allow the psu to switch back on to a ready state.

 

[InvalidError]

How is C6/C7 implemented any differently on 4th-8th gen than on 3rd gen, or 1st gen for that matter?

The difference isn't in the implementation or meaning per se, but in how fast the CPU can go in/out of sleep: prior to Haswell, going in/out of C6 sleep used to be a very slow operation (milliseconds scale) but with Haswell, this can happen thousands of times per second (microsecond scale) and to achieve power savings from going in/out of deep-sleep ~100X faster, the VRM and the PSU by extension, need to be able to cope with those much faster, bigger and more frequent transients.

Prior to Haswell, there wasn't an explicit power specification on how quickly the PSU had to react to load changes. Starting with Haswell which introduced those much faster C6 transitions, Intel's power spec now requires that PSUs be able to cope with load transients of 8A/10us IIRC. That's sleep to full power or full power to sleep in 10us for a 96W CPU. If the PSU can't keep up with that while C6 is enabled, the CPU crashes, which is why many people have to disable C6 sleep to make their Haswell and newer PCs stable when using lower quality and/or older PSUs.

 

[Darkbreeze]

Because they don't handle it differently. They handle it exactly the same. With different values. Just to plug on numbers, if a 3rd gen has a working range where absolute low cpu voltage is 1v, 4th gen absolute low voltage would be something like 0.3v. A group regulated psu might have a 0.8v threshold, a dc-dc 0.2v threshold. They work exactly the same for C states, S states and P states, the difference being the cpu voltage values. At C0-C3 you are still looking at normal cpu voltages, but at C4 the cpu lowers its requirements to maintain clocks as the clocks are not on. That trickle voltage on a Haswell cpu is just lower than Ivy, dropping down to 0.7v at C4. That's plenty for a dc-dc, but going below the group regulated threshold.
The actual process is the same, but being analog, the group regulated psu just can't remain in an operating ready state if the cpu voltage drops too low. There's simply not enough voltage in the cpu to allow the psu to switch back on to a ready state.

That makes no sense to me. The spec (C6, C7, S5, etc.) determines the values. Any changes to those values would either have to be approved by the ACPI governing body or be written into the UEFI, making it a motherboard focused change rather than one based on the CPU architecture.

I mean, maybe you're correct, but considering there is no mention in any of the data sheets regarding changes to the specification for any of the core-i generations, and no mention on any of the ACPI specification listings regarding value changes, I don't see it being accurate.

You'd think that if there was a change to the value of a specification, it would require the creation of a new specification OR a new category of specifications. Otherwise, you could not claim adherence to the specification. I don't need an explanation of how or why the group regulated designs don't work with the specific states, I already fully understand that. My only concern is to why this is not a concern for older gen's that specifically say they adhere to the same exact (Not counting the C7 state for some of the older models, although most of them after the 2nd gen DO say they support that as well) specification but don't apparently have the same issues.

It does not seem accurate to say than X processor uses X amount of power while in C6 state but Y processor uses something different.

 

[Darkbreeze]

How is C6/C7 implemented any differently on 4th-8th gen than on 3rd gen, or 1st gen for that matter?

The difference isn't in the implementation or meaning per se, but in how fast the CPU can go in/out of sleep: prior to Haswell, going in/out of C6 sleep used to be a very slow operation (milliseconds scale) but with Haswell, this can happen thousands of times per second (microsecond scale) and to achieve power savings from going in/out of deep-sleep ~100X faster, the VRM and the PSU by extension, need to be able to cope with those much faster, bigger and more frequent transients.

Prior to Haswell, there wasn't an explicit power specification on how quickly the PSU had to react to load changes. Starting with Haswell which introduced those much faster C6 transitions, Intel's power spec now requires that PSUs be able to cope with load transients of 8A/10us IIRC. That's sleep to full power or full power to sleep in 10us for a 96W CPU. If the PSU can't keep up with that while C6 is enabled, the CPU crashes, which is why many people have to disable C6 sleep to make their Haswell and newer PCs stable when using lower quality and/or older PSUs.

So, then the situation is that newer Core-i processors have a different transition specification which requires that it react to conditions requiring it to resume to full power state, and that the problem, unlike what is advertised or reported by practically every tech site and power supply documentation, it is not actually whether or not the CPU supports those states, but instead is actually that they support an entirely different criteria that probably ought to be it's own specification rather than be lumped in with the generalization that it's a C6/C7 support issue? Since those older architectures DO support those states, as they as written in the ACPI specifications.

 

[InvalidError]

it is not actually whether or not the CPU supports those states, but instead is actually that they support an entirely different criteria that probably ought to be it's own specification rather than be lumped in with the generalization that it's a C6/C7 support issue?

Newer CPUs come with new VRMs and new power specs for the different CPU states all the time. I wouldn't consider C6/C7 any differently. CPUs got much faster at coming in/out of C6 and now PSUs need to cope with that new reality just like MoBo manufacturers do with their VRMs.

BTW, just had a look at Intel's 2013 PSU design guide and the transient rate is 1A/us, 200mA/us worse than I remembered.

 

[Darkbreeze]

Ok. I'm grasping the concept a little better now. But it still seems as though the idea that the problem with group regulation is being advertised inaccurately when it's broadcast that C6/C7 support is the defining factor rather than referencing it as a resumptive characteristic exclusive from whether or not the actual capability to adhere to the state of low power defined by the spec. It CAN reduce power to that level and it CAN resume back to full power operation, therefore IS C state compliant. CANNOT however do it at the RATE defined by Intel so perhaps should be defined separately, ie, C6b/C7b.

I dunno. Seems to fly in the face of traditional hardware standards, where if a new capability exists then a new standard is defined and components are allowed to be either compliant or not, rather than "compliant but in a new faster way" so we're changing what we consider compliant to that spec. Usually, that's what defines a NEW standard is a newer faster version of the old standard. LOL.

 

[InvalidError]

I dunno. Seems to fly in the face of traditional hardware standards, where if a new capability exists then a new standard is defined

C6/C7 is merely a CPU state, it isn't a 'standard' of any shape or form, every CPU has a different internal implementation of it with different characteristics where the CPU's VRM does the bulk of the work. The only difference with Haswell is the sudden increase in PSU performance expectations, which is why Intel published a new ATX 1.31 PSU design guide in 2013. That's your new spec or standard to meet.

 

[goldstone77]

^just a lot of great information I wish we had a like button!

 

[Darkbreeze]

100% true. That's why I asked him specifically for some clarification since I couldn't seem to get any elsewhere. Thank you for that by the way. Much appreciated.

 

[ko888]

Intel's Haswell and later processors feature a C7 sleep state that isn't compatible with all power supplies. Haswell and later processor's low power sleep state power draw is substantially lower than that of previous generations, and it can trigger some PSUs' over-voltage or under-voltage protection circuit forcing the need for a PSU reset.

According to Intel's IDF (Intel Developer Forum), Haswell and later processors can enter a sleep state called C7 that can drop the processor VRM's +12V current draw to as low as 0.05 Amp. Even if the sleeping CPU is the only load on the +12V rail, most power supplies can handle a load this low. The potential problem may arise when there is still a substantial load on the power supply's minor rails (i.e. the +3.3V and +5V). If the load on these minor rails are above a certain threshold (which varies by PSU brand and model), the +12V can exceed ATX12V specs (voltages greater than +12.6V) or the voltage on the +5V rail may drop below the 4.75V minimum allowed by the ATX12V specs. If the +12V and/or +5V rail is out of spec when the motherboard comes out of the sleep state, the PSU's OVP (Over-Voltage Protection) and/or UVP (Under-Voltage Protection) circuit may get triggered preventing the PSU from running and will cause the power supply to "latch off". This will require the user to cycle the AC power switch on the back of the PSU to reset the PSU's protection circuit.

This problem may occur with PSUs with a secondary side that uses the group regulated circuit design due to the +12V and +5V rails sharing the same choke. PSUs with their secondary side that utilize independent (indy) regulation or DC-to-DC circuit design don't suffer from this problem.

Prior processor generations to Haswell had a minimum sleep power consumption of 6 Watts. With Haswell the minimum sleep power consumption was lowered to one-tenth of that at 0.6 Watt.

To truly meet Intel's own Haswell PSU testing methodology none of the PSU's output rails should ever go outside of ATX12V specs under any load condition.

Anyone building a Haswell or later system from scratch are better off getting a compatible unit to begin with.

Anyone building a Haswell or later system using their existing group regulated PSU should disable processor C6/C7 power state support in their motherboard's BIOS setup if they encounter the PSU latching off when resuming from sleep state.

 

[Karadjgne]

I love it when ko888 finally has had enough and pipes up. His answers are Techy enough to support the rationale, and in English enough to be understood by me. (Still looking for the Best Answer button....)

 

[Darkbreeze]

Ok, so apparently, contrary to almost every piece of documentation out there, it is NOT the C6/C7 states that are the culprit in this issue, which is what is commonly passed around even in most the major tech articles covering this subject, but instead is only the C7 state. Looking more closely at the data sheets I see that ko888 is correct in his assessment that it is the C7 state, alone, that is responsible for the problem because prior to 4th gen Intel processors there was only C6, with no C7 at all. So it is C7 only that is the cause of the problem.

This actually makes sense to me now. So, beyond this, I'm wondering why it is after Haswell there has been little to no continued conversation or revisitation of the fact that ALL 5th, 6th, 7th and 8th gen Intel Core-i processors also implement the C7 state and therefore all have the same issue with group regulated power supplies? I've assumed this is true, despite it sort of fading away quietly, but it seems as though this is something that ought to be kept in the forefront of PSU conversations with these processors since group regulated designs still don't seem to be going away.

I've actually heard people arguing that NO, that's only on Haswell processors and had to argue the point until it became senseless to do so anymore. Seems like the power supply manufacturers also ought to make changes to their designations to reflect that this "certification" applies to these newer models as well, not just Haswell. Regardless that we mostly know this, the average user hasn't a clue other than the drivel they see via the marketing speak.