AMD's RX 7900 Series Suffering A0 Stepping Issues and Releasing Unfinished RDNA 3 Silicon; Possible RDNA 3 Revisions Coming 2023

[leadeater]

10 hours ago, CommanderAlex said:

The original problem whether AMD used A0 stepping speculation doesn't appear to be true, although AMD has not replied to inquiries made by Tom's Hardware yet. 

Just like to point out that using A0 stepping silicon isn't actually a problem at all. If it's ready then it's ready, be it first go round or one hundredth.

 

Pretty sure in the Tom's article they point out that Nvidia on multiple occasions have used A0 silicon in final products, we should be giving pats on the back for being able to do this and saying job well done.

[porina]

On at A0 thing, I saw this. I don't know the source but their methodology seems to be extracting this information from posted vbios, so there may be some variations from earliest shipped product and this list.

 

As an example of things not going well, there is Sapphire Rapids. Intel's still delayed server CPU range. They're on more than a handful of respins now but I can't find a list I saw previously.

 

 

[AlTech]

When AMD releases A0 silicon: Why is AMD releasing unfinished hardware with fatal flaws??!!! REEEEEEEEEE....

 

When AMD releases A1 silicon: Why is AMD taking so damn long to release their products??!!!!

 

When Nvidia releases A0 silicon: OMG this is the best thing since sliced bread! Must buy immediately!!!

 

When Nvidia releases A1 silicon: OMG they refined it and optimized it even more! Must buy!

[porina]

37 minutes ago, AluminiumTech said:

When AMD releases A0 silicon: Why is AMD releasing unfinished hardware with fatal flaws??!!! REEEEEEEEEE....

Before this incident I don't think anyone cared what steppings were used, without going back to perhaps Core 2 era when I recall there was one every overclocker was after.

[Forbidden Wafer]

6 hours ago, porina said:

As an example of things not going well, there is Sapphire Rapids. Intel's still delayed server CPU range. They're on more than a handful of respins now but I can't find a list I saw previously.

I think it is the one from Toms Hardware:
 

Quote

There are also many examples of chips that had issues in the design/verification process that required multiple steppings to come to market. For instance, Sapphire Rapids was last known to be on the 12th stepping, and it still hasn't shipped in volume (A0, A1, B0, C0, C1, C2, D0, E0, E2, E3, E4, and E5 stepping — technically 5 base spins). Naturally, that has led to severe production delays and missed launch dates.  

Source: https://www.tomshardware.com/news/amd-addresses-controversy-rdna-3-shader-pre-fetching-works-fine

[starsmine]

1 hour ago, porina said:

Before this incident I don't think anyone cared what steppings were used, without going back to perhaps Core 2 era when I recall there was one every overclocker was after.

Sandy Bridge p68 chipset (the chipset. not the cpu) had a massive hardware bug that created degraded sata performance over time. 
People went out of their way to make sure they did not get an A0/B0 (I think C0 fixed it? I dont rememeber) boards.

But yes. the stepping is a nothing burger other then stating A1/B0 could fix some of the issues on a refresh next year that we already knew was coming for some parts. No one ever used to care.

[BiG StroOnZ]

I'm wondering if this is something that has to do with the stepping (very interesting behavior regarding clocks on Navi 31):

 

 

[CommanderAlex]

4 minutes ago, BiG StroOnZ said:

I'm wondering if this is something that has to do with the stepping (very interesting behavior regarding clocks on Navi 31):

 

 

I watched his video this morning and it appeared that the card is being power limited, only peaking to 397W. His overclocking methodology was just terrible...my $0.02. 

[BiG StroOnZ]

9 minutes ago, CommanderAlex said:

His overclocking methodology was just terrible...my $0.02. 

 

Is there something he could have done differently? 

[CommanderAlex]

3 minutes ago, BiG StroOnZ said:

 

Is there something he could have done differently? 

Undervolt and overclock, gradually overclock VRAM and front clock instead of just focusing on front clock. 

[BiG StroOnZ]

15 minutes ago, CommanderAlex said:

Undervolt and overclock, gradually overclock VRAM and front clock instead of just focusing on front clock. 

 

From what he explained in the video, you can't overclock VRAM and front clock at the same time. Is that incorrect? I imagine he would have been able to figure that out if it was possible.

[CommanderAlex]

41 minutes ago, BiG StroOnZ said:

 

From what he explained in the video, you can't overclock VRAM and front clock at the same time. Is that incorrect? I imagine he would have been able to figure that out if it was possible.

At 3:53 of the video, he did say you cannot auto OC GPU and auto OC VRAM at the same time. 

 

[BiG StroOnZ]

52 minutes ago, CommanderAlex said:

At 3:53 of the video, he did say you cannot auto OC GPU and auto OC VRAM at the same time. 

 

 

Correct, but from what I can tell in the Adrenaline UI from the video, the manual tuning only shows the core able to be adjusted; while underneath where it says VRAM tuning, it says disabled. Is enabling VRAM tuning, as simple as clicking that slider where it says disabled and it enables it? If you can still overclock the VRAM, he just didn't demonstrate it, I stand corrected. And also RDNA 3 GPUs are chiplet based (at least Navi 31 and 32), so I thought that might be a new thing with the MCDs. I haven't used Radeon GPUs (ATi) since the late 90's / early 2000's, so be patient with me.

 

From what he was demonstrating in the video though, if as you say is a power limit, overclocking the core results in VRAM drastically downclocking. Also, he didn't try the reverse though (overclock VRAM and see if core downclocks); from what he was actually doing in the video, he seemed to only overclock the core and focused on that alone.
 

Regardless of this, his results showed, even when he was able to get the VRAM to not downclock, and got respectable clock speeds from the core. That still, performance degraded in benchmarks and he got lower FPS compared to when the clocks were lower (as in lower clocks equaled more performance than higher clocks).

 

Now about a week and a half ago, when these NAVI 31 design issues were appearing; people were claiming that because of a strict power consumption limit, Navi 31 suffered from a design flaw that severely limited its peak boost clocks. 

 

Could this behavior that Jay is showing, be exactly what these rumors were talking about? This is why I was wondering if this is related to the stepping rumor.

 

Lastly, the flaw from the flagship was purportedly fixed with Navi 32 and Navi 33. Meaning, perhaps, in a new stepping of Navi 31 (7950 XT, 7950 XTX?), they may be able to address this peak boost clock limitation. If there is one. 

[CommanderAlex]

1 hour ago, BiG StroOnZ said:

Correct, but from what I can tell in the Adrenaline UI from the video, the manual tuning only shows the core able to be adjusted; while underneath where it says VRAM tuning, it says disabled. Is enabling VRAM tuning, as simple as clicking that slider where it says disabled and it enables it? If you can still overclock the VRAM, he just didn't demonstrate it, I stand corrected. And also RDNA 3 GPUs are chiplet based (at least Navi 31 and 32), so I thought that might be a new thing with the MCDs. I haven't used Radeon GPUs (ATi) since the late 90's / early 2000's, so be patient with me.

Hmm, under Manual tuning [Custom], I am able to control VRAM tuning and GPU tuning independently from each other, albeit this is on my 6900XT, a RDNA 2 GPU, not RDNA 3 with it's chiplets so whether you can adjust them or not will be something to investigate with. 

1 hour ago, BiG StroOnZ said:

From what he was demonstrating in the video though, if as you say is a power limit, overclocking the core results in VRAM drastically downclocking. Also, he didn't try the reverse though (overclock VRAM and see if core downclocks); from what he was actually doing in the video, he seemed to only overclock the core and focused on that alone.

That's what appears to be happening, as Jay pointed that out as their card isn't drawing more power and have noticed that VRAM clockspeeds plummet, most likely to reallocate power to the GPU overclocking and reducing power to the VRAM. 

 

1 hour ago, BiG StroOnZ said:

Now about a week and a half ago, when these NAVI 31 design issues were appearing; people were claiming that because of a strict power consumption limit, Navi 31 suffered from a design flaw that severely limited its peak boost clocks. 

I've been reading up into this and some have said that after the events on the 12VHPWR cable with Nvidia and the wattage draw, AMD possibly wanted to keep power consumption down in order to "compete" with Nvidia on that front. Whether or not these claims are true, we'll have to wait and see what the potential of RDNA 3 can do. 

 

1 hour ago, BiG StroOnZ said:

Lastly, the flaw from the flagship was purportedly fixed with Navi 32 and Navi 33. Meaning, perhaps, in a new stepping of Navi 31 (7950 XT, 7950 XTX?), they may be able to address this peak boost clock limitation. If there is one. 

Correct, that's what the sources from the articles I have in my original post point to which are just leakers saying this. I think possibly AMD wanted to stay on the side of safety for right now, just due to the chiplet design and possibly issues related to A0 stepping that hasn't been hammered out yet. Drivers appear to have a big part, as always and AMD's weakest link, will need to be made to improve performance issues we've seen in the review videos where % gains aren't that big compared to a 6950XT. 

[starsmine]

7 hours ago, BiG StroOnZ said:

I'm wondering if this is something that has to do with the stepping (very interesting behavior regarding clocks on Navi 31):

 

 

Considering most other people are getting 10% improvement from overclocking, he is probably doing something very wrong.

They did not do Witcher with reference XTX

Jay clearly does not know how to overclock AMD cards is my conclusion. My best guess is its because he is not undervolting, hence his issue running into power limits.

Why does Jay keep publishing things before running it through say, a forum first to answer his questions >.>
He says he looked around on the 16th... this was on the 15th
https://www.techpowerup.com/review/sapphire-radeon-rx-7900-xtx-nitro/
and this was on the 12th
https://www.techpowerup.com/review/amd-radeon-rx-7900-xtx/

 

Quote

Overclocking the Radeon RX 7900 XTX is complicated. Just like on RDNA2, you can no longer dial in a specific clock frequency, but you're operating with min/max sliders to set a range. Surprisingly, changing the GPU clock either does nothing for performance, or it results in a loss of performance or crashes the card when set too high. I haven't been able to find any setting that actually increases performance.

It's great to see that AMD has unlocked the slider length (in the past you often ran out of slider length, effectively capping the OC potential).

Memory overclocking works well, but you have to monitor performance. Once memory becomes unstable, there will be no visual corruption, but performance goes down.

I found the best gains for overclocking in general can be found when setting the power limit to maximum (+15%) and combining that with an undervolt that reduces heat output and GPU temperature considerably, which lets the boost algorithm boost higher, for longer.

Quote

Overclocking the Radeon RX 7900 XTX is complicated. Just like on RDNA2, you can no longer dial in a specific clock frequency, but you're operating with min/max sliders to set a range. While on the AMD reference cards, changing the GPU clock does nothing for performance, or it results in a loss of performance or crashes the card when set too high, on this 3x 8-pin custom design things are slightly different.

Due to the increased power limit, the GPU frequency sliders can actually help. Just like on the reference card, for maximum OC performance you should increase power limit to max (+15%), then find your stable memory clock. Memory overclocking works well, but you have to monitor performance. Once memory becomes unstable, there will be no visual corruption, but performance goes down.

Once you have those two, start undervolting until your card is no longer stable—do not touch GPU clocks at this point. As you reduce the voltage, you'll see clockspeed go up automagically, because there is more power headroom to do so. But at some point you'll hit a wall with the clocks, they simply do not go higher, even though they still are 50 MHz below the "clock limit" slider. This is normal for RDNA2 / RDNA3, the "maximum" really seems to be "maximum minus 50 MHz."

At this point increase the max clock slider by 100 MHz, leave the minimum clock slider alone.

Now you can see GPU clocks increasing beyond the previous "wall." Keep reducing voltage until your card becomes unstable. If you hit the frequency wall again, increase max clocks by another 100 MHz.

Done!

 

[StDragon]

3 hours ago, CommanderAlex said:

Drivers appear to have a big part, as always and AMD's weakest link...

You could probably throw the source code at ChatGPT for an improvement 

 

Spoiler

sarcasm

 

[Fendrick]

Nothing really new from team red, I swear they come up with things which sound great then decide to handle it like it is some off the cuff hobby project where it don't really matter how it turns out, just enjoying the adventure they had making it.

[CommanderAlex]

7 hours ago, Fendrick said:

Nothing really new from team red, I swear they come up with things which sound great then decide to handle it like it is some off the cuff hobby project where it don't really matter how it turns out, just enjoying the adventure they had making it.

Kinda like the adventures of Fury, Vega/Radeon VII, and HBM. It just died off quickly after the VII and focus on GDDR6 for right now. 

[starsmine]

Jay to his credit fixed a number of issues with his first video. Still tries to make a few weird claims, but whatever, overall an improvement

 

[BiG StroOnZ]

On 12/19/2022 at 11:46 PM, starsmine said:

Considering most other people are getting 10% improvement from overclocking, he is probably doing something very wrong.

 

Spoiler

 

They did not do Witcher with reference XTX

Jay clearly does not know how to overclock AMD cards is my conclusion. My best guess is its because he is not undervolting, hence his issue running into power limits.

Why does Jay keep publishing things before running it through say, a forum first to answer his questions >.>

He says he looked around on the 16th... this was on the 15th
https://www.techpowerup.com/review/sapphire-radeon-rx-7900-xtx-nitro/

and this was on the 12th
https://www.techpowerup.com/review/amd-radeon-rx-7900-xtx/

 

Definitely seems like TPU's methodology is the best here, as many other outlets I looked at were struggling to get 3-5% improvement. Funny though, in KitGuru's review, in their overclocking section, AMD reached out to them and gave the following recommended settings for tweaking the XTX:

 

Quote

Undervolt -125mV, +100MHz memclk and +15% powerlimit

 

So it seems AMD is very aware of the Undervolting benefits and power limitations on these cards. 

 

1 hour ago, starsmine said:

Jay to his credit fixed a number of issues with his first video. Still tries to make a few weird claims, but whatever, overall an improvement

 

Spoiler

 

 

Watched this today and he seems to be getting great results from the Undervolt, although he didn't examine the VRAM overclocking as closely as I would like.

 

Seems TPU has had the best OC results with the 7900 XTX, especially the clocks they are getting compared to Jay and other outlets. Wonder if they are an outlier. Any info on what the community is getting? Haven't checked Reddit or Overclock.net.  

[-rascal-]

On 12/18/2022 at 9:34 AM, starsmine said:

Sandy Bridge p68 chipset (the chipset. not the cpu) had a massive hardware bug that created degraded sata performance over time. 
People went out of their way to make sure they did not get an A0/B0 (I think C0 fixed it? I dont rememeber) boards.

But yes. the stepping is a nothing burger other then stating A1/B0 could fix some of the issues on a refresh next year that we already knew was coming for some parts. No one ever used to care.

 

Yup, but it was the P67 chipset, B3 revision/stepping.

[AnonymousGuy]

I'll tell you how this goes:

 

Executive: we need to have a production release on A0 silicon.

Engineers: not possible

Executive: what's in the way of that

Engineers: a) b) c) d)

Executive: What do you need to fix that

Engineers: time

Executive: you don't have that so how else can you fix it.

Engineers: waiver

Executive: ok then, release the buggy shit.

[CommanderAlex]

10 hours ago, AnonymousGuy said:

I'll tell you how this goes:

 

Executive: we need to have a production release on A0 silicon.

Engineers: not possible

Executive: what's in the way of that

Engineers: a) b) c) d)

Executive: What do you need to fix that

Engineers: time

Executive: you don't have that so how else can you fix it.

Engineers: waiver

Executive: ok then, release the buggy shit.

Then:

 

Executive: Why isn't A0 performing well on certain games??

Engineers: We still need to fix a) b) c) d)

Executive: Looks like you're working thru the holidays to get it fixed!!!

Engineers: Okaay