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

[Infuriare]

Isn't all overclocking technically "unsupported"? Not only do you have APIs for 3rd party utilities to make use of but you also conveniently bake an overclocking utility into your Radeon settings application.
Unless there is concrete proof that HBM overclocking can somehow by some miracle actually, physically damage cards, why has HBM overclocking for Fiji been left in the dark?
At first I thought this was a driver bug that would eventually get resolved, but alas after 8 months it appears to be a deliberate attempt to gimp the card.

Nice one Kaap. Here's a fun fact for all your Fiji overclockers.

Fiji’s MCLK overdrives in discrete steps, so while various overclocking tools are increasing in 5Mhz steps the MCLK is actually only able to support 500.00/545.45/600.00/666.66MHz and right now it just rounds to the nearest step. If you are intending to overclock HBM on Fiji, then 545Mhz works well on all four of my Fury X cards. 600MHZ proved a step too far on each card and resulted in instability.

So for all those people setting their MCLK at 570Mhz, it's actually running at 545Mhz.

https://forums.overclockers.co.uk/posts/29045102/

AMDMatt, 18 months ago did you or did you not reveal that HBM overclocked in straps, but also confirm that you overclocked 4 of your cards to 545mhz? How on earth did you achieve such an amazing feat if this is not "supported"? It's also quite ironic how you literally used the word "support" in your post.

 

[LtMatt]

https://forums.overclockers.co.uk/posts/29045102/

AMDMatt, 18 months ago did you or did you not reveal that HBM overclocked in straps, but also confirm that you overclocked 4 of your cards to 545mhz? How on earth did you achieve such an amazing feat if this is not "supported"? It's also quite ironic how you literally used the word "support" in your post.

I used a third party tool to enable HBM overclocking, i was never able to do it via Wattman as it was, not supported.

 

[Kaapstad]

I used a third party tool to enable HBM overclocking, i was never able to do it via Wattman as it was, not supported.

I did try pointing this out earlier but to no avail.

Wattman did not exist until the RX480 turned up.

 

[gupsterg]

No hard feelings, i know where you are coming from. Sorry it's not the answer you were hoping for.

Thank you for finally acknowledging 'where I'm coming from' .

@Scougar @WeSbO @mmj_uk

Thank you for understanding what I see as an issue , many do not seem to grasp it .

As you saw in the video OverDrive panel which was present in driver from card launch till Wattman came about was mediocre . Yes MSI AB, Trixx, etc allowed some extra elements but not a patch on bios mod TBH .

So pretty much for over a year+ there was no application official or 3rd party which could enable an owner to give voltage to a specific DPM (Power State). OD or a 3rd party app we OC'd highest state (DPM 7), all we had access to as voltage control was 3rd party apps. These had a voltage offset, this is 'globally' applied to all states. So even if idle clock (and/or intermediate states) were not being OC'd, we would gain an increased voltage. VBIOS has 8 DPMs, viewing the PowerTune WhitePaper we can see that PowerTune actually has a ton load of granularity and far more states. By doing bios mod I had perfected 2 VBIOS for my card, a stock clocks undervolted and OC clocks overvolted in required states only and rest undervolted.

I enjoyed over a year where I could move to a new driver and have my card as I wanted without having to set anything in driver or a 3rd party app to apply profile at OS load. I could be in any OS of my choice and card would be as I wanted.

This is just the 'touching the skin' of the finesse VBIOS mod allows.

Many who have failed to understand my gripe, custom VBIOS applying your preferred setup is so hassle free.

Fiji as stated before has far more of a better Fan table in PowerPlay. I opted to tweak the 'Fuzzy Logic' fan profile, which is the factory setup on these cards. Based on 'target temperature' and how SMU monitor powerdraw, etc it will ramp cooling solution as required to meet the temp target. No application so far allows this.

As stated before there are more attributes in the 'Fan Table'. You can actually set a GPU clock MHz which will limit cooling not to ramp up until it passes that GPU MHz, so you can have a quieter card. An example of this in action, SWBF runs sweet on Fiji, I use 1440P with Ultra preset, I cap FPS at ~88FPS as the MG279Q I have has FS range of 35-90Hz. I enable 'Power Efficiency' by default in driver, now as GPU FPS is capped it never really reaches DPM 7 plus PE has an affect in keeping GPU MHz in check more tightly as it's not being loaded hard. I end up with a truly quiet card.

Next it is not all about unlocking what is not officially given or a 3rd party app does not give access to. I have actually protected my card more than AMD/AIB has.

I reduced OCP on both loops of VRM controller, so as I know what clocks it gonna do I do not need the higher limit. I also lowered VRM throttle temps, there are two in the PowerPlay, one for GPU and other HBM. Now if my cooling solution was to give up the ghost whilst gaming or left running f@h unattended, the card would be further protected from damage.

All I can take away is it suited AMD not to block HBM clock changes before, it suits them now.

We have had the posts "its supported under VEGA" so let's go get VEGA now.

I did try pointing this out earlier but to no avail.

Wattman did not exist until the RX480 turned up.

And to no avail you have not got the reason why I have shown the issues by VBIOS mod, as originally in the VEGA thread AMD Matt said it's an issue with 3rd party apps. Which it is not.

You can use 3rd app on any driver from launch of card (~June 15) to Jan 17 and you can OC HBM.

A 3rd party app or data value change of VBIOS which is AMD code has no effect on v17.x.x driver, now it has been ~8 months of driver releases where we have been locked out.

I have a HD 5850 in one of my rigs, I use last good driver for it now, IIRC Omega driver from ~end of 2014. Fiji is not technically obsolete like that and not unsupported by AMD. They have pretty much kicked Fiji owners in the crouch when it really wasn't needed.

 

[Kaapstad]

snip

Mods are not the responsibility of AMD to support in any way shape or form.

NVidia are also very good at locking people out of doing stuff on their cards.

Both AMD and NVidia are in the business of selling cards that will operate inside their normal operating limits, they are not in the business of dealing with stacks of RMAs where people have modded their cards.

 

[gupsterg]

All I have a gripe with is for over a year HBM clock changes were not an issue. Why did AMD let it go on for so long?

Would they not have been getting stacks of RMAs when card was actually in retail channel, from 2017 it was not even in retail IIRC. If it was, no one was really gonna jump on it for the price it was and looking at what was on the market.

In hindsight at the time of buying Fury X I should have gone with the competitor. I know by my purchase decisions it makes not a large difference to xyz company. But I reckon I would have been better off with a 980 Ti. I would have 6GB of VRAM, I'd have got a great OC fix from it, yeah G-Sync would have been an added cost vs FreeSync but at least I'd have had a card with 'longer legs'.

 

[Freddie1980]

All I have a gripe with is for over a year HBM clock changes were not an issue. Why did AMD let it go on for so long?

I've read that big post of yours (#84) where you talk about the BIOS mods you had used to get all this HBM overclocking to work, there has been a change to power delivery with newer cards due to some security requirement that has effected memory overclocking. Steve at Gamers Nexus mentioned this in a Vega video he done earlier this week, it may or may not be relvent (given your talking about Fiji, I don't know how retrospective this new protocol is), I'll see if I can dig out the video for you.

 

[gupsterg]

Getting HBM OC is simple. Newer drivers blocking any gain is beyond me.

The bigger post #84 was to explain how lacking any OS SW is compared to what can be done by VBIOS mod. It was also to show how I protected by card more than what AMD/AIB would do.

The VRM is bigger. Loosely speaking power delivery is same but improved.

The 'security feature' that VEGA has is a processor which checks VBIOS for modification. There is a digital signature that it must reference.

This signature has existed since 7xxx series. All cards before had no hardware to check signature, driver did at OS load. AMD can choose to enforce the driver checking signature at OS load, but it can also be modified out. For a lengthy while Hawaii bios mod could not be done as driver checked signature. You had limited VBIOS which you could use, for example Shammino/Shammy did PT VBIOS (he had it signed AFAIK), then there were The Stilt mining VBIOS (again signed)

Next example when Polaris launched you couldn't use a modified VBIOS, the author of CRU (ToastyX) created an app which patches driver, this patch is based on a mod that Lordkag did for me on Fernando's Win RAID forum. Der8auer/Elmor did release some mod VBIOS for Polaris, again signed.

The digital signature originally came about when UEFI become mainstream IMO, there is a AMD PDF from an event where UEFI.org members discussed what they did to be compliant. UEFI has 'Secure boot' feature secure boot means basically non approved firmware can not be used.

AMD VBIOS (Atombios) is Hybrid VBIOS. It has a Legacy section which motherboards which are non UEFI use. Motherboards that have UEFI and CSM is disabled in mobo UEFI, will use the UEFI/GOP module in VBIOS but as the the GOP does not have all info it uses the Legacy section to setup card.

So as UEFI has secure boot feature the Legacy section must be secured. SO AMD implemented a digital signature in Legacy section. The mobo checks a signature in UEFI/GOP of VBIOS to know if it is modded. The VBIOS GOP module checks digital signature in legacy section to know if it modded. So there were ~2-3 layers of security prior to VEGA. If all is well then card will post in 'pure UEFI' environment (ie CSM: Off SB: On).

If you set mobo to CSM: Off SB: Off you still have 'pure UEFI' environment but mobo will not check digital signature in GOP of VBIOS. The GOP will check signature in Legacy section. The GOP can me modded to not do this.

If your mobo is set to CSM: On (usual default for mobos) and SB is [Auto] or [On] or [Off] it performs no checks to see if VBIOS is modded. In this case AMDs only security was driver checking at OS load digital sig.

VEGA is locked down hard. Pretty much all that went on before can not.

But if the digital signature gets cracked (which has been done before, one mention on a miners forum) then it can get changed by AMD, but open to being cracked again. If this crack occurs I reckon it will be from the mining community, to me seems there are a fair few adept hands in that community, they also have massive GPU power to run algorithms for cracking IMO.

Ahh well good whilst it lasted.