AMD Gimping HBM clock on Fiji after v16.12.2 WHQL driver

[gupsterg]

I came about as a Fiji owner ~ Feb 16, I had been knee deep in Hawaii bios mod and enjoyed it. SomeOCN Fiji owners were after bios mod, I gave pointers on OCN where I usually post. As this did not bare fruit I then got a Fury Tri-X. I picked this as it was on promo at the time and I also wanted to experience shader unlock.

I did gain 3840SP, I then also picked up a Fury X again on promo. In total I have had 1x Fury Tri-X, 8x Fury X and 2x Fury Nitro, info on these exploits is on OCN.

When I joined the Fiji owners club on OCN there were some highlighting that HBM clocks in steps. ie 500 / 545 / 600 and 666.

As I found no info I asked a member for source, I was then pointed to a post by AMD Matt on OCuk, I also found the same info on Guru3D posted by him.

I kept an open mind, I was a disbeliever initially. Using AIDA64 GPGPU bench it was clear that HBM did OC in steps. So then I iterated to as many members as I could the info.

Now for months, since v17.x.x drivers AMD have gimped HBM performance gain from HBM clock increase.

Here is a zip of today testing, link.

No single result on v17.8.1 has beat 1145 / 545 Memory Copy, I have in the past also done this testing on my i5 4690K at 4.9GHz CPU and 4.4GHz cache. Here is link of some testing back in May.

*** edit ***

I know this card very well I have done in the past 250+ 3DM 13 benches on this card besides the other hours and hours of usage for f@h, gaming, tweaking, etc.

This card on any driver below v17.8.1 can not do more than HBM 545MHz, it will artifact at desktop and look WOW on v17.8.1 I have had ~30min usage (compare clock in linked zip for screenie on 1145 700 with below link).

http://i.imgur.com/rGpDuew.jpg

 

[gupsterg]

Why this thread has been started on OCuk has been as AMD Matt is most active on here. I have mentioned his username on OCN several times when the subject of HBM clocks not gaining performance on later drivers and we have had no information or help. Usually a OCN member will get forum/email notification of member being mentioned in xyz thread so it is brought to their attention and participate in thread.

I mean not to agitate any members or upset anyone .

I am only an owner that is posting/sharing as to what I and others are experiencing on Fiji. As I wish to have a resolution or reason given as to why the behavior of HBM clock change has no effect on performance in v17.x.x drivers.

Are we Fiji owners being gimped by AMD?

Is it to force Fiji owners to move on to VEGA?

In the VEGA 64 owners thread another Fiji owner placed a question to AMD Matt, here.

I have quoted the relevant posts here.

Yes I'm aware of that one, we'll investigate.

For now just leave HBM voltage alone, it overclocks by 15-25% without any voltage increases.

Is there any word on fixing drivers for Fiji chips? Drivers later than the January WHQL do not allow application of HBM overclocks.

Sounds like third party tools broke with the introduction of Wattman. Sorry, but we are not going to fix that. Third party apps will need to be updated, HBM overclocking was never enabled in the driver.

So here we have it that 3rd party apps are the cause of HBM clock increase not gaining performance. HBM Overclocking was never enabled in driver.

This is somewhat correct and somewhat not.

Wattman was feature of Polaris Drivers, for months it was not available on past AMD cards. When Crimson ReLive drivers were introduced past AMD cards got Wattman.

Spoiler

So prior to that we had OverDrive.

Now in PowerPlay of VBIOS are OverDrive Limits. On all AMD cards the VRAM is also above what is stock value for VRAM clock on Fiji it was not. OD RAM Limit was 500MHz and as the PowerPlay value for setting HBM clock was also 500MHz no slider appears in OverDrive.

Below is screenshot of OD from March 16.

Spoiler

Now when I change only OD Limit for RAM in PowerPlay of VBIOS or say we used MSI AB to 'Extend OverClocking Limits' we would have HBM Slider active.

Spoiler

So there we have it that driver did display HBM clock slider. I could have a VBIOS on card with only PowerPlay OD limit for RAM extended to what I wish and then increased HBM Clock through driver. I would not need 'MSI AB 'Extend OverClocking Limits'.

What does 'MSI AB 'Extend OverClocking Limits' do?

What this feature does is enable a registry key in Windows to have SoftPowerPlayTable. The key is within AMD drivers of registry.

This PowerPlay table will the take precedent over what is in firmware of VBIOS. This allows users of MSI AB to use driver to set card as they want. When the key is active you don't even need MSI AB to change things TBH, if you are proficient in Registry editing you can edit the AMD SoftPowerPlay key as you would VBIOS, etc.

Next we have PMC25 explaining it was not 3rd party apps issue.

No, that has been debunked by various OC tool developers. It's bugged / switched off at a driver level.

All the third party tools have updated their API paths since Wattman was introduced. Drivers continually break it (like the penultimate driver did to Fiji where GPU core clocks would stick at 300Mhz), but they continually update (as they have now done with the last bug).

They can't fix HBM overclocking because the driver refuses to apply it. This goes for BIOS mods which change the values and don't use 3rd party apps at all.

If this isn't fixed, then people will just wonder if the same thing is going to happen to Vega.

Have you actually enquired about it? I'm beginning to suspect that it was 'bugged' on purpose. Could be that it was to discourage miners from returning cards where they'd fried the HBM with 24/7 use and over ambitious overclocks.

That would not effect miners at all though, as they don't need the latest gaming drivers and all just use the last working driver from January.

At this point I posted in the VEGA owners thread.

I don't use any 3rd party app.

I don't use Wattman or OverDrive.

Any v17.x.x driver gains no performance with HBM clock increase via VBIOS. But v16.12.2 WHQL and earlier do.

In my Fiji bios mod thread on OCN there are several members which have same occurrence. On OCN several members in the Fiji owners club have same occurrence.

There have been several posts by members there where they get obscene HBM clocks on v17.x.x drivers and no gain. The moment they go to v16.12.2 WHQL they see those clocks not achievable but they get a performance gain.

As far as I and many OCN Fiji owners are concerned AMD have gimped Fiji driver from v17.x.x onwards.

Shame really. I have highlighted to you several times and mentioned your OCN username in posts to make you aware of situation.

Then there is the issue of SWBF, that mesh quality must be lowered on v17.x.x driver. Otherwise game artifacts. And it does this even on stock VBIOS, W7 and W10 clean installs. And as you are on AMD Community site and participating in thread there you'll already be aware it's not just a single Fiji owner with issue. v16.12.2 WHQL does not have this issue either, stock or OC, W7 or W10.

It seems to me AMD FineWine is getting corked these days .

I luv'd Hawaii, I owned 4 cards.

Fiji at first I was not gobsmacked, slowly slowly it really grew on me.

Buildzoid was interested in what bios mod could do for him on Fiji. As he had killed a Hawaii card trying to hard mod some elements, there is post on TPU/his site about it. So I really got the Fury to do mods as others who I highlighted info were not succeeding. Then it just grew on me.

I grabbed a Fury X next, on promo. Worked out cheaper than say value of Hawaii with WC block. I sold all Hawaii cards without loss, I went on to get another 7 Fury X and 2 Fury Nitro. On OCN there is thread about 'ASIC Quality' where I uploaded data, supporting info that The Stilt had given on Hawaii in this context.

Fiji IMO gained very little over the course of drivers. And now we are left with gimped drivers. In hindsight I should have ket Hawaii and gone WC.

I think AMD FineWine has been 'Hyped'. I never kept abreast of VR or have experienced it. I recall lots of threads/posts about a feature which eas supported on Polaris but not Fiji. Dunno if it got addressed. As a flagship card it really hasn't had much luv from AMD on driver front IMO.

I have waited for VEGA and pricing is deplorable IMO. I invested in FreeSync and now feel if I jump to nVidia I'll be seeing a double whammy on swap out.

For me real shame about AMD on GPU front for past few years as an enthusiast.

I don't often visit oc.net and if i do, i rarely login.

No performance increase from HBM would be expected, as it's not intended to be overclocked in the drivers. It may have been possible to do it via third party apps before, but that's not our job to fix it because it stopped working when using those apps.

We didn't intend for HBM to be overclockable on Fij, this functionality was never exposed in Overdrive/Wattman. Sorry nothing i can do here.

Matt did you miss the part where I said I don't use third party app?

I OC'd HBM using VBIOS. I can show you AIDA64 GPGPU and 3DM FS gaining performance with HBM clock increase on v16.12.2 WHQL (plus every earlier driver) and not on any v17.x.x driver.

This same method was how I proved to members that HBM1 clocked in steps. As many and indeed I contested this concept you posted originally on OCuk, Guru3D and other places.

Come on mate, do you think we support bios modifications? Furthermore, do you think that we will use valuable engineering resources investigating how to enable HBM overclocking on a product that was never designed to support HBM overclocking?

It was enabled before. AMD switched it OFF. There is a difference. That's tacit support being purposely removed.

Almost no Fijis weren't capable of 545, and a lot were capable of 600. Both offered large boosts, particularly when OC'ing the core.

This gimps performance.

I still haven't heard of any HBM failing under normal gaming / productivity use with overclocks. The only chance of damaging it comes with over volting it - which only miners (who use the old, working driver) and people messing around with LN2 did.

HBM overclocking was never exposed in Overdrive or Wattman.

Yes technically HBM overclocking was not exposed in OverDrive, as stated and shown before this was down to how OverDrive Limit for HBM was not above HBM clock in PowerPlay. Once this was made like how other cards have increased OverDrive VRAM MHz Limit I had slider for HBM just like any other AMD card.

When WattMan was introduced a slider would not appear. So somewhat AMD took something away from driver when ReLive edition was released, but we still had performance gain from HBM clock increase for several months. IIRC Relive was introduced ~June 16, the final driver that gives HBM performance gain with clock increase was released Jan 17.

Next I reply to AMD Matt again.

I am not asking or stating is bios mod supported.

I am stating that where we saw performance gain from HBM clock increase on every driver until v17.x.x why has it been gimped now?

I added nothing to VBIOS to make HBM OC, I only change PowerPlay HBM clock from 500MHz to what I want. This is what a AIB would do when selling a OC edition card. Which is also similar to what a SW tool would be doing via driver/OS.

I can see it is futile discussing this with you Matt, you are bouncy the question back as an AMD rep and not seeing it as an end user POV. Your initial attempts were to bounce this at 3rd party apps, now you are digging at bios mod.

I think I'll do some benching and postings on various sites, so perhaps prospective AMD VEGA owners may know that a year or so down the line a flagship card gets gimped by AMD just as nVidia has been highlighted in media.

Big difference, HBM2 overclocking is supported on Vega in our drivers, it was never supported on Fiji. If it was supported, in the bios or via drivers, it would've been available in Overdrive/Wattman.

If HBM2 overclocking ever stops working on Vega, please let me know and I'll ensure engineering takes a look at it.

Let's get back on topic now as instructed by Stanners.

Then I have asked AMD Matt to join discussion in this thread.

He has not so far.

As I noted he was online and posting VEGA info I asked again politely.

Thank you, I'll be looking out for a VEGA deal .

There is link to a thread in this post, any chance of giving your time , thanks .

I don't have anything more to add from what i said earlier, so there is no point getting into a debate about it.

You are avoiding the subject , post what you did here over there, let me show you what I do and how the driver is blocking gains from HBM clock increase .

There is plenty of other info I can share how AMD can manipulate driver as they want .

Give me a fair crack of the whip instead of upsetting members here by posting in this thread.

What we have here IMO is AMD have placed a block within driver.

My card will not do 700MHz HBM (ie 1400MHz effective speed) on any driver prior to v17.x.x, but with any v17.x.x driver it will and I still get performance as card is at 500MHz. Reference zip in 1st post of thread and a linked screenie at end of post 1.

There is no doubt in my mind that AMD have gimped Fiji.

There is no AMD FineWine for Fiji.

I have noted over the course of ownership ~5% gains at best in performance with drivers. Was I after more? not really. I just wanted bug fixes and things not taken away that have been there.

SWBF has been an issue of Fiji for months as well. Not only experienced by me but by others as well, link. I have religiously used AMD driver feedback page since every v17.x.x driver to highlight issue with SWBF. Lowering Mesh Quality is a workaround and not a fix, this workaround works on all v17.x.x drivers but on v16.12.2 WHQL I do not need this workaround.

 

[LeMson]

Maybe it's because of the wonkyness of card builds that's everywhere now?

 

[Martini1991]

It's a bit iffy to say they don't support bios mods when AMD brought about dual BIOS's on the GPU'S which is direct benefit of those who would bios MOD

 

[Deleted User 234324]

No surprise really, they need to make Vega look good even if it isn't.

 

[Andy]

If it looks like a fish, swims like a fish and is a fish, it will probably smell like a fish too. Disgraceful if true.

 

[gupsterg]

Maybe it's because of the wonkyness of card builds that's everywhere now?

Unfortunately I don't think this. I have had too many AMD cards. Some Fiji did 545MHz without any need for HBM voltage increase. I had one sample that did 600MHz a lot better than others. To me 545MHz about the average potential. Not massive but at least something.

Fiji has this issue of losing GPU performance scaling with voltage increase. You can gain a small consistent gain in performance even at stock clocks with under volting. This is not because GPU was hitting powerlimit I have thread on OCN where I did 250+ 3DM 13 benches and used same PL for all runs.

So HBM clocking gained me back this scaling issue from voltage increase.

It's a bit iffy to say they don't support bios mods when AMD brought about dual BIOS's on the GPU'S which is direct benefit of those who would bios MOD

Bios mod support I'm not contesting on an official POV. Even when we use Wattman/OverDrive we see a warning that overclocking is basically warranty void IIRC. So even if we have an official app support OC it really isn't that we are covered in a water tight fashion.

We are basically sold 'stuff' with a premise of OC/tweaking to a degree.

No surprise really, they need to make Vega look good even if it isn't.

I do some what think AMD have not focused on Fiji well. There were countless posts on web concerning how a VR feature was not enabled for Fiji but for Polaris IIRC. As I have never used VR or read much I do not have the info but will search for it.

At all readers.

VEGA having the extra 'Security feature' for bios mod to satisfy Microsoft is a load of bull IMO as well.

Here I will quote a reply I did on OCN to a member to on my view of this (link to original post).

I'm thinking part of windows secure boot requirements is to blame.

I really don't think it is part of this TBH.

Let's ignore what VEGA has/does for a moment. Let's talk about Fiji and below.

'pure UEFI mode' is achieved by setting CSM to Off/Disabled in UEFI of motherboard.

CSM = Compatibility Support Module, now if a GPU does not have UEFI/GOP module in VBIOS it is not usable with CSM: Off/Disabled.

When CSM is off with Secure Boot Off and VBIOS has UEFI/GOP module, the motherboard UEFI is not concerned with validating UEFI/GOP module. The UEFI/GOP module in VBIOS won't validate itself either. The UEFI/GOP module must validate Legacy section of VBIOS as it needs that info to make card function. The UEFI/GOP module references the same digital signature that drivers do at OS load . So here we have a feature to secure UEFI mode.

Next when CSM is off with Secure Boot On and VBIOS has UEFI/GOP module, the motherboard UEFI will validate UEFI/GOP module in VBIOS. The UEFI/GOP module in VBIOS has a digital signature for itself. So now the process is motherboard validates VBIOS UEFI/GOP module, VBIOS UEFI/GOP module validates Legacy section of VBIOS. So we have two features to secure UEFI mode.

IMO all is covered on a MS POV for secure UEFI post mode.

So as you can see we had pretty much 3 features to secure VBIOS.

i) Digital signature in Legacy Section of VBIOS, which OS driver refs to know it is unmodified. Which AMD can decide to have on or off.

ii) When CSM: Off, SB: Off, UEFI/GOP module of VBIOS validates Legacy section.

iii) When CSM: Off, SB: On, mobo UEFI validates VBIOS UEFI/GOP module, VBIOS UEFI/GOP module validates Legacy section.

So on VEGA we have all of above, plus a security processor checking VBIOS to make sure it has not been modified.

You will recall in Fiji Bios mod there was discussion on how a modified VBIOS when using CSM: Off black screens/does not post. How this was resolved was by using a custom UEFI/GOP module in VBIOS. This modification made UEFI/GOP module not validate digital signature in Legacy Section of VBIOS. We just can not use Secure Boot as the signature in UEFI/GOP module is not updated to reflect changes within it. As you can guess when Secure Boot is on the mobo UEFI must get validated as well, so it can't be modified not to ignore the signature in UEFI/GOP module of VBIOS.

So there were several layers of security already to make UEFI org/MS happy.

Now on VEGA due to this security processor we can't :-

i) modify UEFI/GOP module not to ignore digital signature in Legacy section so we can have 'pure UEFI mode' with CSM: Off.

ii) modify Legacy section to have changes we want.

Also AMD have disabled i2c comms on VEGA FE for sure by VBIOS, RX VEGA seems as if it is the same, so we can't do certain things as we did on cards before.

Thanks for the links. :thumb:

REP +1

I read about 4-5 pages of the Preliminary view of AMD VEGA Frontier Edition Bios thread. No wonder we haven't seen much of you. you've been busy.

Any chance of fooling the security by showing it official Bios, while injecting/replacing data table bios queries to modded bios or expanding PCI-E control through the crossfire connection?

No idea on your suggestions.

The methods that are successful so far are pretty much similar IMO. The member that has made mods on Linux has done so by loading a modified copy of bios when OS loads. The WinOS registry trick is pretty much the same, the driver gives priority to copy of PowerPlay in registry when it exist at OS load.

If it looks like a fish, swims like a fish and is a fish, it will probably smell like a fish too. Disgraceful if true.

I agree .

It is true for me and several other owners .

 

[gupsterg]

Now just as added info, all the 8 Fury X I have owned never had a UEFI/GOP module in VBIOS. I purchased Fury X cards between ~Feb 16 to late 2016. AIBs etc never really made Fury X / Nano have UEFI compatible VBIOS 'out of box' either. Fury Tri-X/Nitro I had received with UEFI/GOP module within VBIOS 'out of box'. The Fury Tri-X was bought ~Feb 16 and my first Fiji card.

I had noted some AIBs such as ASUS and Gigabyte had updated VBIOS with UEFI/GOP modules on their websites. So an owner could update VBIOS and use 'pure UEFI' mode (ie CSM: Off in mobo UEFI). AMD did release a updated Fury X and Nano VBIOS, link.

No flash tool in package, potentially warranty void as well, as a card bought from AIB they would state that the updated AMD VBIOS was not from them.

There are several debacles on Fiji IMO, which may have gone unnoticed to most people IMO.

 

[LeMson]

Unfortunately I don't think this. I have had too many AMD cards. Some Fiji did 545MHz without any need for HBM voltage increase. I had one sample that did 600MHz a lot better than others. To me 545MHz about the average potential. Not massive but at least something.

Fiji has this issue of losing GPU performance scaling with voltage increase. You can gain a small consistent gain in performance even at stock clocks with under volting. This is not because GPU was hitting powerlimit I have thread on OCN where I did 250+ 3DM 13 benches and used same PL for all runs.

So HBM clocking gained me back this scaling issue from voltage increase.

No, literally wonky, with differing heights between cpu and hmb

 

[gupsterg]

Ahh yes VEGA has this issue . Fiji did not AFAIK.

 

[Freddie1980]

No, literally wonky, with differing heights between cpu and hmb

His talking about HBM1 I think your talking about HBM2 which is reportedly 0.4 of a mm shorter then the GPU.

 

[ubersonic]

No, literally wonky, with differing heights between cpu and hmb

If that really bothers you I would avoid a Threadripper or Ryzen CPU, or any of the old school Intel dual cores. Whenever you have multiple dies on the same package there will always be tiny varations in height. They don't actually affect anything as the TIM evens it out.

 

[LeMson]

Ow sorry wrong video card lolz

 

[ubersonic]

There is no doubt in my mind that AMD have gimped Fiji.

Well maybe you should read what AMDmatt and others have told you, or maybe just read your own posts as you actually explain it all yourself. HMB overclocking on Fiji was never supported, only made available via 3rd party apps, those apps are not compatible with the newer drivers and so are unable to make it available, stop having a go at AMD and start badgering the creators of those apps to make them compatible with the modern drivers.

 

[subbytna]

If it looks like a fish, swims like a fish and is a fish, it will probably smell like a fish too. Disgraceful if true.

Nvidia have been doing this for ages, all about forcing you to upgrade. Now Red is doing it as well as green, it seems.

 

[gupsterg]

Well maybe you should read what AMDmatt and others have told you, or maybe just read your own posts as you actually explain it all yourself. HMB overclocking on Fiji was never supported, only made available via 3rd party apps, those apps are not compatible with the newer drivers and so are unable to make it available, stop having a go at AMD and start badgering the creators of those apps to make them compatible with the modern drivers.

You have missed the part where I have stated I don't use 3rd party apps.

As shown in 2nd post when PowerPlay of VBIOS OverDrive Limit is made to be not equal to stock of HBM clock (which is how OD limit is on all other cards) then you had slider for HBM in OverDrive in drivers prior to Relive edition.

You have also missed the part that the driver is not giving performance from clock increase on v17.x.x driver when same VBIOS is used between it and v16.12.2 WHQL. What does that tell you? Let me say it clearly the clock change by VBIOS or by any 3rd party app is being blocked on v17.x.x driver IMO.

AMD can do this IMO.

Let me give you another few examples how the driver can do things which you may not know.

AMD enforced driver to check signature in Legacy section of VBIOS in Hawaii, so when you modified VBIOS it would not get to OS. Later when this was removed you could modify VBIOS. Polaris also had that this.

You may recall Polaris ref PCI had a balancing issue on VRM, where power was being drawn at x Watts from PCI-E slot, the fix was to implement a bias to have the portion of VRM supplied by PCI-E plugs to have power loading balance.

You have missed the point on what I am badgering about. You too would want the performance back which is being blocked by a newer driver if you had the issue.

 

[Scougar]

Not supporting memory overclock? That is pretty ******* lame regardless of whether anyone thinks AMD can 'do whatever they want'.

This is pretty much akin to Creative removing features from the soundblaster drivers.

 

[gupsterg]

Not supporting memory overclock? That is pretty ******* lame regardless of whether anyone thinks AMD can 'do whatever they want'.

This is pretty much akin to Creative removing features from the soundblaster drivers.

Exactly .

I like doing VBIOS mod so card is as I want. ie as if it came from factory as I wanted it. I then can install any driver and know the card is at clocks, fan profile, power limit, etc as I want. This was my reason and others in ploughing hours into Hawaii bios mod in 2015. From that I learnt a lot.

Roll on get Fiji, sweet I thought I have also HBM voltage control via VBIOS. So from early 2016 til late 2016, me and other Fiji bios modders are luv'in it. Come 2017 drivers and we have been blocked from performance gain on HBM clock increase.

Not cool IMO.

People who have not done bios mod or aware of how it is done do not get the finesse it can have.

Prior to Wattman being supported on Hawaii, Grenada, Fiji, etc, all we had was 3rd party apps to do voltage control. All the apps did a global voltage offset for all DPM states. With bios mod we could modify each state clock and VID. We had so much control IMO. And as stated you have zero issues with drivers as each driver just picks up settings with VBIOS and does what it needs to.

If someone takes the time to view AMD VBIOS they will see it has a structure, this structure is pretty much universal. In the legacy section we have command and data tables. The command tables between like GPUs did not change at all, so take a 290X Matrix card and a bog standard ref PCB and they have same command tables. What changes are data tables.

These data tables contain values for say clocks, etc or VRM differences. These tables have structural formats and data values. So all that happens is in a way is data values are changed for x card vs y. It is not a total different VBIOS as sorts, loosely speaking.

All I did on Hawaii and Fiji is pretty much change a value of say GPU/VRAM clock as I wanted. If a PowerLimit was too low for said clocks and would have needed a PL increase in driver I change the value in VBIOS and driver think 0% PL is new data value.

The driver from v17.x.x is blocking HBM clock changes by VBIOS or 3rd party app.

Next lets talk about fan profile.

VBIOS on Fiji has real sweet fan profile section, it has far more data values than Hawaii. DO we get access in OverDrive? NO. Did we get access in Wattman to full values? NO.

If they are not going to make SW apps have what we want and people modify VBIOS to have what they want and it was allowed for months and months and to block it when GPU isn't even in retail channel is totally illogical in my POV.

 

[WeSbO]

To make sure I've understood correctly. So when you use a modded bios on Fiji, with oc'ed HBM, those HBM speeds are not taken into account on new drivers, at it runs at stock HBM speeds ?

 

[Zeed]

Welcome Ryzen Master. Well with **** AMD pulled with vega this is just cherry on top of cake...