AMD Bristol/Stoney Ridge Thread

[The Stilt]

Does Bristol Ridge even have DDR3 support in its IMC? I'm thinking not.

It does, for obvious reasons.
Technically one could make a AM4 board with DDR3 support, but that would require custom code (AGESA) and also violate the infrastructure specification.
I don't think that anyone would be interested with a AM4 motherboard, which would only support Excavator based CPUs anyway.

 

[nathanddrews]

I haven't checked RAM prices in a while, but last I looked, DDR4 was cheaper per GB than DDR3, which had also taken a bit of a nosedive. 32GB for $120 or something like that. The cheapskate in me would like to see DDR3 support, but overall I'd rather see DDR3 abandoned in favor of pushing DDR4 performance higher and higher.

 

[KTE]

4+2 (VDD_CR / VDD_SoC) will be the most common config. 3-phase "NB" (VDD_SoC) config is extremely rare and supported just by a single controller model (AFAIK). Gigabyte seems to be using three phase VDD_SoC so that they can skimp in the component quality. Those 4+3 phase Gigabyte boards appear to be using ISL95712, which is a jellybean part (the very reason they are using it). Four integrated gate drivers and analog outside the communications (SVI2).

The proper, digital VRM controllers (IR / Infineon) support >= 2 secondary output phases. Those can be doubled or even quadrupled, but it is definitely not necessary as long as the board is designed properly and high quality components are used. You can design a 2-phase circuit which can supply < 120A of constant current, so the number of the secondary phases supported by the controller is the least of the issues. However as always, quality comes at cost...

Quality > quantity, with no exceptions made.

For a 8C/16T Zeppelin I wouldn't even consider a motherboard with a 4+2 phase config.

Why not ?

Excluding OC...

With Agena/Deneb/Thuban, plenty of 4+2 provided the same or similar OC/user overall experience, as any expensive DFI/Asus/Gigabyte 6+2. So why would someone pay much more?

Sent from HTC 10
(Opinions are own)

 

[The Stilt]

Why not what?

 

[DrMrLordX]

It does, for obvious reasons.

Oh, I had forgotten that Carrizo supported DDR3. Oops, my bad.

Why not ?

Excluding OC...

With Agena/Deneb/Thuban, plenty of 4+2 provided the same or similar OC/user overall experience, as any expensive DFI/Asus/Gigabyte 6+2. So why would someone pay much more?

My impression is that the stress on VRMs is going to be directly related to the current draw of the CPU. You can get a basic idea of how much current a CPU is going to pull based on its vcore (or in the case of CPUs with multiple voltage planes, its voltage per plane) rated against its power draw in W. A 95W Deneb or Thuban (think 1055T) with an operating voltage if 1.3-1.4v is going to pull less power as current than a 95W Summit Ridge with what is probably going to be a voltage in the 1.1-1.2v territory, if not lower. The more current draw you have, the more VRM phases you want on the CPU side.

Hence the need for 8+2 (or 4+1 w/ doubler) as a minimum for something like Summit Ridge.

 

[KTE]

Why not what?

Why not your last quoted line.

Sent from HTC 10
(Opinions are own)

 

[The Stilt]

Why not your last quoted line.

Sent from HTC 10
(Opinions are own)

Because a 95W AM4 CPU has similar TDC to a ~ 155W 45nm AM3 CPU and EDC even higher than that. Even the properly built mainstream 1151 boards usually have 4+2 phase VRM or better. And they are intended to be used mainstream parts with TDP ranging from 35 to 65W, much lower than on Zeppelins.

The VRM requirements depend on the amount of current drawn. Because of that the VRM requirements for e.g a 95W TDP 45nm part running at 1.45V are significantly lower than the requirements for a 95W TDP 14nm part running at 0.975V.

 

[VirtualLarry]

The VRM requirements depend on the amount of current drawn. Because of that the VRM requirements for e.g a 95W TDP 45nm part running at 1.45V are significantly lower than the requirements for a 95W TDP 14nm part running at 0.975V.

Good point!

 

[KTE]

Because a 95W AM4 CPU has similar TDC to a ~ 155W 45nm AM3 CPU and EDC even higher than that. Even the properly built mainstream 1151 boards usually have 4+2 phase VRM or better. And they are intended to be used mainstream parts with TDP ranging from 35 to 65W, much lower than on Zeppelins.

The VRM requirements depend on the amount of current drawn. Because of that the VRM requirements for e.g a 95W TDP 45nm part running at 1.45V are significantly lower than the requirements for a 95W TDP 14nm part running at 0.975V.

I understand that but I'm trying to understand why you say they're not able to run Zen. What exactly limits this on a 4+2 or is it just pushing it?

You're saying that you don't expect any typical 4+2s to:

a) Run a 95W Zenb at all?
b) Run it for long?
c) Run it reliably without damaging the CPU?
d) Run it without limiting/decreasing the clocks?
e) be able to OC it?

Which of the above, and if any other reasons?

Sent from HTC 10
(Opinions are own)

 

[The Stilt]

Same exact issues as with the existing cheap AM3+ and FM2+ boards, when 125W or 95W TDP parts are used.

It is matter of reliability, efficiency and performance, mostly.
The CPU cannot reach or sustain it's full performance since the VRM protections will kick-in and force the CPU to reduce it's speed (VRHOT throttling). In the past when the controllers didn't have as advanced protections, the VRM turned into a puff of smoke and usually claimed the CPU as a collateral.

All VRM controllers which are compatible with AM4 have the sufficient protections, so it is unlikely that a excess VRM stress / overheating would cause any permanent damage.
Same way as on the previous AMD platforms (AM3+ & FM2+).

Like I said before, it is entirely possible to build a 3 or 4 phase VRM which is more than sufficient to support a 8C/16T Zeppelin running under full load 24/7. However the thing is that the high-end components you need for such configuration to work are extremely expensive, in relation to the cheaper ones. You can usually build a 16+ phase VRM featuring lower quality components for the same price as a top quality 3-4 phase VRM. If you think that any of the "usual manufacturers" will use even mainstream quality components in their cheap boards, you're pretty delusional. If mainstream quality components would be assigned as "Grade B", the components the cheapest boards from these "usual manufacturers" feature are somewhere between "Grade D" and "Grade G". On top of that they take additional shortcuts, like creating doubled VRM without the proper hardware (connecting phases in parallel, overloading the drivers), skimp in filtering, etc. These people have no issues in lowering the MTBF expectancy of the product by 20% if they can save two cents on each of them. You need to remember that these boards are mostly designed and manufactured in a country where selling fake eggs, concrete filled wallnuts, rice made of plastic and gutter oil are also a common practice. It's harsh and slightly exaggerated, but gives you the idea.

 

[Abwx]

In the past when the controllers didn't have as advanced protections, the VRM turned into a puff of smoke and usually claimed the CPU as a collateral.

It has often nothing to do with the VRMs capabilities but rather to the inductances being saturated and unable to store more energy, due to this saturation the excess power is entirely dissipated by the VRMs wich are not cooled enough to take such power amount, and they would be that generaly their junction/case thermal resistance is too high, rendering useless an increasement in VRMs coolers size..

Fortunately some progress has been made and inductances that can deal with these kind of current are common and very cheap such that a 4 phase PSU has no trouble supplying over 100W, just look at reference Polaris card PSU capability with 5 phases...

 

[The Stilt]

It has often nothing to do with the VRMs capabilities but rather to the inductances being saturated and unable to store more energy, due to this saturation the excess power is entirely dissipated by the VRMs wich are not cooled enough to take such power amount, and they would be that generaly their junction/case thermal resistance is too high, rendering useless an increasement in VRMs coolers size..

Fortunately some progress has been made and inductances that can deal with these kind of current are common and very cheap such that a 4 phase PSU has no trouble supplying over 100W, just look at reference Polaris card PSU capability with 5 phases...

VRMs failing due the excess temperatures has nothing to do with core saturation. Besides, it is pretty hard to sature the cores on a inductor, which has 40-50A Isat rating at higher temperature than they are ever operating at. You can find several pictures of mainstream AM3+ boards (ASRock, Gigabyte, MSI) on which the VRM has run so hot that it has caused discoloration and delamination of the PCB. Some of them have been no doubt user errors, however there are certain models which are prone to failure due their VRM structure. And none of these boards are the bottom bin ones, I might add.

AMD reference designs since Cayman have always been built like tanks. The Ellesmere (RX 480) has TDP of 110W, which includes several other high-current planes besides just the core. Yet the reference board is using a fully digital, native six phase solution for the GPU power plane. IR3567 digital controller, CHL8510 gate drivers and HS & LS from Magna. In order to find anything even remotely similar configuration on a motherboard, you'll need to look at the flagship motherboards. That's because high-end components are expensive, as I said earlier. AMD doesn't skimp on their reference cards.

 

[Concerned Citizen]

So when are retail nice chips going to be available?
"Q1 2017" when, I want a date .PS:And numbers.

 

[Abwx]

VRMs failing due the excess temperatures has nothing to do with core saturation. Besides, it is pretty hard to sature the cores on a inductor, which has 40-50A Isat rating at higher temperature than they are ever operating at.

What matters is the peak current and this latter is 40% higher than the average current, a core can saturate on very short and repetitive periods and this will slowly heat up the VRMs without causing their destruction, but with the board aesthetic results you re quoting as a consequence, untill they eventually ultimately fail for good.

At 100A a single phase out of 4 will drain 2.5A from the 12V rail, that s this current that is switched by the a pair of VRMs, if they heat up it means that the inductance core is losing efficency, be it by core saturation or increased copper resistance due to temp, or both.

 

[KTE]

That's the problem.

To date, I've never seen stock validated CPUs unable to function fully and properly to an end user (including Turbo Boost) with the average VRM builds.

The stock VRMs only became a problem for OCers.

Some, like the entry-level Sapphire builds for Phenom (PI-AM2RS780G), were better OCers than mainstream and higher end boards MSI K9A2 Platinum/most 770/890GX. Whether this was due to better 4+1 VRMs (Panasonic caps, ferrite chokes and Infineon FETs IIRC) or other factors, wasn't ever ascertained tho.

The problems with OC that came from VRM design were more isolated to individual components... as Max Icc being demanded was at the maximum Idsat range for a FET (17-25A FET 4-phase), and current runaway at high temps/currents causes the severe degradation/instability/blowing of such FETs.

If I look at it in that sense, what you say makes sense with the amperage now being needed with Zen. That would relegate most low-end 4-phase VRM MBs out of compatibility (anything lower than Idsat 25A 120C FETs and good quality rest).

But then, how will they validate those CPUs for their MBs?

Sent from HTC 10
(Opinions are own)

 

[The Stilt]

What matters is the peak current and this latter is 40% higher than the average current, a core can saturate on very short and repetitive periods and this will slowly heat up the VRMs without causing their destruction, but with the board aesthetic results you re quoting as a consequence, untill they eventually ultimately fail for good.

At 100A a single phase out of 4 will drain 2.5A from the 12V rail, that s this current that is switched by the a pair of VRMs, if they heat up it means that the inductance core is losing efficency, be it by core saturation or increased copper resistance due to temp, or both.

I would highly suggest you check what the controller datasheets or design guide (for any modern controller) has to say about the current balance. CHIL / IR controller (SVI, SVI2, VR12, VR12.5 compliant models) datasheets probably have the most detailed illustrations and descriptions about that. I took you for a much younger person, but apparently you're not

 

[Abwx]

The problems with OC that came from VRM design were more isolated to individual components... as Max Icc being demanded was at the maximum Idsat range for a FET (17-25A FET 4-phase), and current runaway at high temps/currents causes the severe degradation/instability/blowing of such FETs.

If I look at it in that sense, what you say makes sense with the amperage now being needed with Zen. That would relegate most low-end 4-phase VRM MBs out of compatibility (anything lower than Idsat 25A 120C FETs and good quality rest).

Forget about the first decade mosfets, this is what is used in AMD s Polaris :

http://www.infineon.com/dgdl/irf6718l2pbf.pdf?fileId=5546d462533600a4015355ed25bd1a88

There s 5 phase in a reference 480 IIRC, and 200W average power is really no problem if we look at the overclocking tests done by some users, that s 40W average per phase, so even a 3 phase PSU could supply a 100W CPU if design is done right...

 

[The Stilt]

The MSI 5-phase "RX 480" you are probably referring to was actually a RX 470 card (press pictures). The actual MSI RX 480 (with six phases) can be seen in TPU review.

 

[Abwx]

The MSI 5-phase "RX 480" you are probably referring to was actually a RX 470 card (press pictures). The actual MSI RX 480 (with six phases) can be seen in TPU review.

Granted that it s possible but still, 4 phases is reasonable for a CPU that dont get over 95W under Prime 95.

FTR here in France the cheap 4+1 95W Asrock AM3+ of recent design were quite successfull and many people even used them with FX8350, it s just that those ones do not overclock and wont even test it with Prime 95, after all they are not less cautious than whom buy a 125W MB to overclock an FX above 4.4GHz and Prime 95 it...

 

[DrMrLordX]

To date, I've never seen stock validated CPUs unable to function fully and properly to an end user (including Turbo Boost) with the average VRM builds.

The stock VRMs only became a problem for OCers.

I don't suppose you ever owned an AM3 motherboard, did you? I had an MSI 790FX-GD70. I had so very often contemplated flashing the BIOS to support a 1055T so I could run that, but I never got around to it because I already bricked one of those under warranty doing a BIOS flash that should have worked (and they weren't nice enough to give me a replacement board with the latest BIOS when they replaced the brick). Had I been successful, I could have discovered the joy of exploding VRMs that other GD70 users discovered when trying to run Thuban on one of those things.

It was not safe, sometimes not even at stock speeds.

 

[KTE]

I don't suppose you ever owned an AM3 motherboard, did you? I had an MSI 790FX-GD70. I had so very often contemplated flashing the BIOS to support a 1055T so I could run that, but I never got around to it because I already bricked one of those under warranty doing a BIOS flash that should have worked (and they weren't nice enough to give me a replacement board with the latest BIOS when they replaced the brick). Had I been successful, I could have discovered the joy of exploding VRMs that other GD70 users discovered when trying to run Thuban on one of those things.

It was not safe, sometimes not even at stock speeds.

I've tested/owned majority of the 740/760/780/790/880/890 FX/GX/G/V boards. Some times, more than 3 of each rev (did that for MFGs).

Nothing after the first BD tho (left the field completely, sold my desktops/workstations).

What I'm saying is for the average board that fully supports the CPU -- rather than upgrade/update drop-ins which have always had issues.

Sent from HTC 10
(Opinions are own)

 

[KTE]

Granted that it s possible but still, 4 phases is reasonable for a CPU that dont get over 95W under Prime 95.

FTR here in France the cheap 4+1 95W Asrock AM3+ of recent design were quite successfull and many people even used them with FX8350, it s just that those ones do not overclock and wont even test it with Prime 95, after all they are not less cautious than whom buy a 125W MB to overclock an FX above 4.4GHz and Prime 95 it...

I think he means more of "95W Zen has much higher current requirements than previous CPUs, which most 4 phase VRMs cannot sustain" than "95W CPUs can't run on 4+2 VRM MBs".

Sent from HTC 10
(Opinions are own)

 

[Soulkeeper]

What about something like the ir3550, any reason not to want these ?
Am I the only one that likes the efficiency/heat benefits that these single package solutions seem to provide ?

 

[The Stilt]

What about something like the ir3550, any reason not to want these ?
Am I the only one that likes the efficiency/heat benefits that these single package solutions seem to provide ?

PowIRStages from IR (now Infineon) have been and still are the best thing there is. HS, LS and a gate driver in a same package, extremely efficient.
As I said before, the issue isn't that it would be impossible to make sufficient and high quality VRMs with a low phase count (i.e 4+2 phase). It is extremely easy when you use high quality components, such as the IR3550.
The issue is that the manufacturers don't build high quality VRMs in their entry-level or even mainstream boards due the cost of the high quality components.

In case of the IR3550, which would allow 60A current delivery per phase (on paper) the cost ratio is > 6:1.

In case of a 4+2 VRM:

IR CHL8150 digital gate driver = 0.455$
OnSemi NTMFS4C09N HS fet = 0.15341$
OnSemi NTMFS4C06N LS fet = 0.14802$

Total for six phases (4+2) = ~ 4.54$

vs.

IR 3550M PowIRStage (HS, LS, GD) = 4.55$

Total for six phases (4+2) = 27.3$

The components I used in the conventional example are very high quality too, so you can easily achieve much higher costs savings when you start digging from the bottom bin manufacturers catalogs (which for certain manufacturers is a rule, rather than an exception).

 

[hojnikb]

PowIRStages from IR (now Infineon) have been and still are the best thing there is. HS, LS and a gate driver in a same package, extremely efficient.
As I said before, the issue isn't that it would be impossible to make sufficient and high quality VRMs with a low phase count (i.e 4+2 phase). It is extremely easy when you use high quality components, such as the IR3550.
The issue is that the manufacturers don't build high quality VRMs in their entry-level or even mainstream boards due the cost of the high quality components.

In case of the IR3550, which would allow 60A current delivery per phase (on paper) the cost ratio is > 6:1.

In case of a 4+2 VRM:

IR CHL8150 digital gate driver = 0.455$
OnSemi NTMFS4C09N HS fet = 0.15341$
OnSemi NTMFS4C06N LS fet = 0.14802$

Total for six phases (4+2) = ~ 4.54$

vs.

IR 3550M PowIRStage (HS, LS, GD) = 4.55$

Total for six phases (4+2) = 27.3$

The components I used in the conventional example are very high quality too, so you can easily achieve much higher costs savings when you start digging from the bottom bin manufacturers catalogs (which for certain manufacturers is a rule, rather than an exception).

Woah, thats a huge difference in pricing. No wonder manufacturers skimp here.

Hmm, what about FIVR... Does that make boards vrm any simpler ? I would guess so, looking at the haswell boards.