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

[CommanderAlex]

 

 

Summary

AMD's flagship Navi 31 GPUs on the A0 stepping appears to have some hardware issues that are affecting performance. This is believed to be the cause for Navi 31's delay by 1 year, resulting in unfinished silicon and a shader prefetch problem. Power issues are exhibited in different programs where the GPU clock varies from as low as 2.2GHz to as high as 2.9GHz. Early RDNA 3 will appear to have a non-working shader prefetch problem, resulting in software patches that hope to correct this until newer revised RDNA 3 silicon can be released next year.

 

Quotes

Quote

As discovered by Kepler_L2, it looks like early RDNA 3 silicon had a non-working shader prefetch HW. This was featured in three chips, the GFX1100 (Navi 31), GFX1102 (Navi 33), & GFX1103 (the APU lineup consisting of Phoenix chips). Now based on the latest GitHub submission. Kepler is also suggesting that Navi 32 GPUs are actually based on the 'GFX1103' IP whereas Navi 33 chips feature the 'GFX1102' IP. The same issues are found on every other chip besides Navi 32 which is going to be featured in several mainstream discrete GPUs on desktop and mobility platforms by early 2023.

 

Going off of this, this can result in months long delays until this can be fixed either partially by software through AMD's Adrenalin drivers or completely new silicon. 

 

Quote

According to Kepler, this is something that cannot be fixed or revised in a few weeks and will take several months if AMD even plans on fixing this. It would be a major blow if newer silicon with a fix comes out a few months later because the majority of gamers would already have bought the early revision of the Radeon RX 7900 series with the unfinished silicon. What's likely is that AMD may prepare a refresh a year later that would address these issues but for now, the Radeon RX 7900 series may have to depend massively on driver-level optimizations to address the unfinished GPU nature of their top RDNA 3 silicon.

 

As a result, users of new RDNA 3 GPUs are reporting strange GPU clock variations across different games/software, resulting in GPU clock speeds as low as 2.2GHz to as high as 2.9GHz. They are also stating that they're being treated as guinea pigs for AMD to launch a product this unrefined.

 

My thoughts

 

I'm pretty shocked that AMD would release RDNA 3 knowing there are issues with the silicon. AMD didn't want to come out of the gate and say that there were issues with performance and instead chose to say "up to 1.7x more performance" in certain games and it definitely shows in the slew of release videos showcasing the 7900XTX and 7900XT. I think if you are on the fence on buying one of these GPUs, it's best to wait or even skip this generation entirely until a refreshed RDNA 3 comes out or RDNA 4 (if that's coming next). 

Sources

AMD accused of treating consumers as 'guinea pigs' by shipping unfinished RX 7900 GPUs | TechRadar

AMD's flagship RX 7900 XTX GPUs discovered to integrate A0 silicon revision affected by serious hardware issues - NotebookCheck.net News

Rumors Allege AMD Shipped Unfinished RDNA 3 Navi 31 A0 GPU Silicon In Radeon RX 7900 Series (wccftech.com)

[xAcid9]

33 minutes ago, CommanderAlex said:

Now based on the latest GitHub submission.

 

July 11 submission is latest?

 

Judging from weird power consumption and leaked rumor about silicon problem before release day, I think they're onto something.

[CommanderAlex]

1 minute ago, xAcid9 said:

 

July 11 submission is latest?

It appears AMD have been having this problems with RDNA 3 since then...and apparently cannot resolve these issues since. 

[porina]

Hardware problems are nothing new, and even if we take this at face value, they still seem to give decent performance. Assuming it requires a hardware fix, this could lead to an interesting refresh down the line.

 

I fell for a trap at launch. I was so focused on what it didn't do, I didn't see what it could. Regardless if this bug exists or not, you're still getting the performance it is offering today.

[Levent]

I have been saying it for a while now but, AMDs graphics card division and software division is in need of major change.

 

It is baffling to something like this get to production. AMD seriously needs to improve the quality of their products, be it hardware or software.

[xAcid9]

25 minutes ago, CommanderAlex said:

It appears AMD have been having this problems with RDNA 3 since then...and apparently cannot resolve these issues since. 

Yeah. The entry still not change since August. 

 

Also if you check XTX BIOSes from TechPowerUp, you can see the naming have A0 in it for reference and XFX Merc card.

 

AMD Reference

 

XFX Merc

 

Asus TUF

 

But how do you actually check the chip naming? GFX1100, GFX1102, GFX1103

[Dietrichw]

I wonder how much it is effecting performance. If it is 5% or less I don’t care, but if it is more then I would be wanting to know when the revision comes. I was about to buy a 7900 xtx that came is stock 30 minutes ago but decide to wait and get more information on this. I may just get a used last gen card and upgrade to the revised hardware if it is significant. 

[AlTech]

According to MLID, AMD didn't want to release RDNA3 this year. Nvidia forced them to by launching 40 series this year.

2 hours ago, Levent said:

I have been saying it for a while now but, AMDs graphics card division and software division is in need of major change.

Management at the GPU Hardware division of AMD has changed. Not sure if a change is needed in the rank and file of the hardware division.

 

Their software division on the other hand needs serious help.

 

With respect to RDNA3, MLID's sources indicated that the performance level the 7900XTX was targeting was not communicated properly in AMD and ultimately AMD as a whole were surprised to see 7900XTX perform the way it did. They expected it to be as fast as they claimed in their presentation it would be, 50-70% faster than a 6950XT.

 

2 hours ago, Levent said:

It is baffling to something like this get to production. AMD seriously needs to improve the quality of their products, be it hardware or software.

I suspect people at AMD had to know about this issue but ultimately concluded that they couldn't fix it in time for release.

 

AMD didn't want to release RDNA3 this year according to MLID. They only released it this year cos Nvidia released the 40 series this year. If Nvidia hadn't released it then AMD wouldn't have released RDNA3 this year.

[AlTech]

5 minutes ago, Dietrichw said:

I wonder how much it is effecting performance. If it is 5% or less I don’t care, but if it is more then I would be wanting to know when the revision comes. I was about to buy a 7900 xtx that came is stock 30 minutes ago but decide to wait and get more information on this. I may just get a used last gen card and upgrade to the revised hardware if it is significant. 

If you're happy with either 6800XT or 6900XT performance and you can find a 6800XT for $500 ish or a 6900XT for $700 then I would suggest grabbing that instead. If not, waiting this out is an option.

[xAcid9]

3 minutes ago, AluminiumTech said:

AMD didn't want to release RDNA3 this year according to MLID. They only released it this year cos Nvidia released the 40 series this year. If Nvidia hadn't released it then AMD wouldn't have released RDNA3 this year.

WOW.. MLID read my mind. I think i should make a YT video as well... but i don't have any anonymous source..

[porina]

18 minutes ago, AluminiumTech said:

AMD didn't want to release RDNA3 this year according to MLID. They only released it this year cos Nvidia released the 40 series this year. If Nvidia hadn't released it then AMD wouldn't have released RDNA3 this year.

Source notwithstanding, I'm not sure I buy that and struggling to process that.

 

Generation launch dates:

May 2016 Pascal
Aug 2017 Vega
Sep 2018 Turing
Jul 2019 RDNA
Sep 2020 Ampere
Nov 2020 RDNA2
Oct 2022 Ada
Dec 2022 RDNA3
 

Nvidia have been on a more or less 2 year cycle though these generations. AMD are more variable. While past performance is no guarantee of future behaviour, if AMD really did want to have a product going against nvidia close to launch then planning for an nvidia 2 year-ish cycle would seem to be the way to go. Or did they expect delays from nvidia for some reason? I'm not aware of a technical reason, and the only reasons are more woolly like macroeconomic conditions and shifting the old products, which I'd argue would be less clear back when AMD would have had to make those decisions. Is AMD not confident enough to release their product when it is ready, as opposed to being "forced" by a competitor doing business as normal?

 

The only thing that makes gives me support this claim has any truth is the FSR 3 announcement. The way it was given and the timing of it just made it sound like they panicked and felt they had to match nvidia.

[Kilrah]

48 minutes ago, AluminiumTech said:

According to MLID, AMD didn't want to release RDNA3 this year. Nvidia forced them to by launching 40 series this year.

Nvidia's schedule has been stable with a release every 2 years for the past decade or so, not like they would have been surprised by it... so that's a BS take. 

[Fasterthannothing]

I don't understand who signed off on this. Absolutely zero excuse to release something broken when they know it's broken.

[CommanderAlex]

1 hour ago, AluminiumTech said:

According to MLID, AMD didn't want to release RDNA3 this year. Nvidia forced them to by launching 40 series this year.

 

This prefetch problem is a major issue

Management at the GPU Hardware division of AMD has changed. Not sure if a change is needed in the rank and file of the hardware division.

 

Their software division on the other hand needs serious help.

 

With respect to RDNA3, MLID's sources indicated that the performance level the 7900XTX was targeting was not communicated properly in AMD and ultimately AMD as a whole were surprised to see 7900XTX perform the way it did. They expected it to be as fast as they claimed in their presentation it would be, 50-70% faster than a 6950XT.

 

I suspect people at AMD had to know about this issue but ultimately concluded that they couldn't fix it in time for release.

 

AMD didn't want to release RDNA3 this year according to MLID. They only released it this year cos Nvidia released the 40 series this year. If Nvidia hadn't released it then AMD wouldn't have released RDNA3 this year.

That's what I've been seeing too from MLID on his coverage the other day on RDNA 3. 

[AlTech]

1 hour ago, porina said:

Source notwithstanding, I'm not sure I buy that and struggling to process that.

 

Generation launch dates:

May 2016 Pascal
Aug 2017 Vega
Sep 2018 Turing
Jul 2019 RDNA
Sep 2020 Ampere
Nov 2020 RDNA2
Oct 2022 Ada
Dec 2022 RDNA3
 

Nvidia have been on a more or less 2 year cycle though these generations. AMD are more variable. While past performance is no guarantee of future behaviour, if AMD really did want to have a product going against nvidia close to launch then planning for an nvidia 2 year-ish cycle would seem to be the way to go.

They wanted to release RDNA3 in 2023. Once Nvidia announced Ada Lovelace AMD couldn't help but launch early compared to their plan to launch in 2023.

1 hour ago, porina said:

Or did they expect delays from nvidia for some reason? I'm not aware of a technical reason, and the only reasons are more woolly like macroeconomic conditions and shifting the old products, which I'd argue would be less clear back when AMD would have had to make those decisions. Is AMD not confident enough to release their product when it is ready, as opposed to being "forced" by a competitor doing business as normal?

I think part of it is they wanted to be sure that they could offload all their RX 6000 parts before RX 7000.

 

Another part is that they may have wanted more time to push 7900XTX harder in clockspeeds and power.

 

It would make sense if AMD expected Nvidia to be delayed for one reason or another but obviously Nvidia weren't delayed.

1 hour ago, porina said:

The only thing that makes gives me support this claim has any truth is the FSR 3 announcement. The way it was given and the timing of it just made it sound like they panicked and felt they had to match nvidia.

 

 

[HenrySalayne]

16 hours ago, AluminiumTech said:

According to MLID, AMD didn't want to release RDNA3 this year. Nvidia forced them to by launching 40 series this year.

It takes month in prep work to get a new GPU on the shelves. AMD launching RDNA3 is not a reaction to Nvidia. In comparison to Intel, AMD is relying on board partners to make graphics cards and they cannot afford to produce 10 million units and put them in a warehouse.

[Doobeedoo]

I was curious about this the first time we heard it. Now as I'm in no rush and wanted to see custom cards anyway, I didn't get one. So revised chip could get a decent uplift it seems, especially if rumored much higher clocks are true. Hopefully new revision doesn't take too long. Few months won't be a big deal, year would be. 

Same how I wait to see new X3D chips which is a good idea.

 

Some extra info:

AMD Addresses Controversy: RDNA 3 Shader Pre-Fetching Works Fine | Tom's Hardware (tomshardware.com)

[porina]

2 hours ago, Doobeedoo said:

Some extra info:

AMD Addresses Controversy: RDNA 3 Shader Pre-Fetching Works Fine | Tom's Hardware (tomshardware.com)

To expand on that, basically shipping A0 means everything worked well enough to ship product. It's not a bad thing. That's not to say there wont be potential for future improvement. This puts the claim that AMD didn't want to ship until next year into further doubt. Maybe that was based on a contingency AMD had in reserve in case further work was indeed needed to get a shipping product.

 

 

[Dietrichw]

It is good know it doesn’t mean much. I guess I am back in the hunt for a card then. 

[starsmine]

On 12/16/2022 at 10:51 AM, AluminiumTech said:

According to MLID, AMD didn't want to release RDNA3 this year. Nvidia forced them to by launching 40 series this year.

that wasnt MLID's take on it. I have went back and watched his previous videos. 

MLID never said AMD was forced to launch this year because Nvidia.

 

Just that they announced launching it this year, CONSIDIDERED delaying it a few months back and decided it was not worth the reputational damage to delay to next year for the reasons of

Delays look bad to investors
Card works well enough
Ironing out all the bugs could take anywhere from 2 months to 6 months if at all.

RDNA 1 stock is low, would be handing over market and mind share to nvidia.

Bringing up that there was talks and consideration does not mean AMD did not want to release it this year, just that it was considered. 

This is a big reason why you have to be careful with leaks, and why companies dislike leaks. Because internal talks of considering things gets misunderstood as a want/desire/promise to people on the outside when its not.

Its not that leaks are false, its that things change.

[AlTech]

5 hours ago, starsmine said:

that wasnt MLID's take on it. I have went back and watched his previous videos. 

MLID never said AMD was forced to launch this year because Nvidia.

Pretty sure he did previously although his latest video today walked that back. I guess that's what you watched.

5 hours ago, starsmine said:

Just that they announced launching it this year, CONSIDIDERED delaying it a few months back and decided it was not worth the reputational damage to delay to next year for the reasons of

Delays look bad to investors
Card works well enough
Ironing out all the bugs could take anywhere from 2 months to 6 months if at all.

RDNA 1 stock is low, would be handing over market and mind share to nvidia.

*RDNA2

5 hours ago, starsmine said:

Bringing up that there was talks and consideration does not mean AMD did not want to release it this year, just that it was considered. 

This is a big reason why you have to be careful with leaks, and why companies dislike leaks. Because internal talks of considering things gets misunderstood as a want/desire/promise to people on the outside when its not.

 

 

5 hours ago, starsmine said:

Its not that leaks are false, its that things change.

Agreed.

[CommanderAlex]

UPDATE: On Tom's Hardware, an AMD spokesperson has made a statement in regard to the shader prefetch problem that has been going around on news outlets. 

Quote

Like previous hardware generations, shader pre-fetching is supported on RDNA 3 as per [gitlab link(opens in new tab)]. The code in question controls an experimental function which was not targeted for inclusion in these products and will not be enabled in this generation of product. This is a common industry practice to include experimental features to enable exploration and tuning for deployment in a future product generation.

The code that was cited in the original posts by Kepler_L2, whom @xAcid9 pointed out to me being months old,  (whom is quoted in the original post) was pertaining to an experimental function that was not intended for final RDNA 3 products, therefore is disabled for now. 

 

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. 

 

My thoughts

This appears to be just growing pains that we've uncovered...yet unless AMD makes a statement in regards to the A0 stepping for clarification whether "unfinished" silicon was used to launch the 7900XTX/XT. I just want to keep you guys updated on where this story is going. 

 

AMD Addresses Controversy: RDNA 3 Shader Pre-Fetching Works Fine | Tom's Hardware (tomshardware.com)

[starsmine]

1 hour ago, AluminiumTech said:

Pretty sure he did previously although his latest video today walked that back. I guess that's what you watched.

No I watched his previous two videos first before commenting that. If he said it ealier then that, I missed it.

[xAcid9]

9 hours ago, porina said:

To expand on that, basically shipping A0 means everything worked well enough to ship product. It's not a bad thing. That's not to say there wont be potential for future improvement. This puts the claim that AMD didn't want to ship until next year into further doubt. Maybe that was based on a contingency AMD had in reserve in case further work was indeed needed to get a shipping product.

 

-snip-

I'd 100% agree with his statement if 7900 series don't have weird power consumption problem. AMD need to fix it a.s.a.p with driver/BIOS update.

 

What? MLID was wrong? Again? Oh my god... *sweep under the rug*

He rarely wrong. 

[starsmine]

1 hour ago, xAcid9 said:

I'd 100% agree with his statement if 7900 series don't have weird power consumption problem. AMD need to fix it a.s.a.p with driver/BIOS update.

 

What? MLID was wrong? Again? Oh my god... *sweep under the rug*

He rarely wrong. 

MLID was wrong how? (in this case)
Not that he cant get his analysis wrong, he has, but saying AMD had internal talks about delaying to next year is not wrong. Saying that his sources at AMD expected x number two months ago is not him being wrong. You can only do analysis on the data you have, and sometimes the data is wrong. That does not make the analysis wrong, the analysis would still be correct based on the data.

Simplified logic. 

2+2 = 4

but you go, "but the initial conditions were 2+3, so it wasn't 4, you were wrong..."

like no, the data was wrong, the analysis is still correct.

And even then, it could very well HAVE been 2+2, 3 months ago, but things changed because things were not finalized yet, so was the data actually wrong? no.

AMD went, our simulations show this, our targets show this. product is made, oh shit it does not match the simulations or targets. BUT, its also still a generational leap. Its still good enough, people buying day 1 enthusiast hardware don't care about idle power much. 

There are talks about RDNA3+ apparently existing, whatever that means, dont know, not enough has leaked out. Hard to analize, could just be mobile, could be more.
They have stated some of the known pipeline issues in RDNA3 have already been fixed for future chips, meaning RDNA4 for sure, does that mean 3+ again, dont know.

7970 HD was A0, 7970 GHZ edition was B0. 

They may also be banking on drivers allowing the cards to be rebenchmarked/"refreshed" when the full Navi 31 cards come out with the 7950 xtx next year, and being closer to their performance targets for the cards.