AMD to Reduce RDNA 4 "Navi 44" Chip Package Size

[OneSorcerer]

Please release a 8600xt performing like a 7700xt with 16gb <= 215mm for my next itx build.

 

[Bad]

A waste of silicon...

RDNA4 would be just another trash tech unless they open up the firmware and allow the end user to tweak the cards to their liking.

As bad as RDNA3 was, the complete lack of MorePowerTool support was the last nail in its coffin... at least for me personally.

I briefly considered a 7800XT but was absolutely appalled by its undervolting behavior. While you can apply an undervolt, the card would just want to hit its power limiter and will use the extra headroom to squeeze out a few megahertz worth of frequency - which would be next to worthless - instead of saving power.

After all, the V/F curve doesn't scale linearly.

And to add insult to the injury, the power limiter on most 7800XTs only goes down by like 8% or so! On a 250W+ card! At least allow us to set the power limiter from -50% to +25%, which was the case with RDNA2.

Suffice to say, I ended up buying a 4070S and it can run at 130-140W all day long with just ~10% performance sacrifice and barely any heat and noise. I ran some benchmarks, and the performance is basically on par with the 320W RTX3080.

 

[Dr. Dro]

A waste of silicon...

RDNA4 would be just another trash tech unless they open up the firmware and allow the end user to tweak the cards to their liking.

As bad as RDNA3 was, the complete lack of MorePowerTool support was the last nail in its coffin... at least for me personally.

I briefly considered a 7800XT but was absolutely appalled by its undervolting behavior. While you can apply an undervolt, the card would just want to hit its power limiter and will use the extra headroom to squeeze out a few megahertz worth of frequency - which would be next to worthless - instead of saving power.

After all, the V/F curve doesn't scale linearly.

And to add insult to the injury, the power limiter on most 7800XTs only goes down by like 8% or so! On a 250W+ card! At least allow us to set the power limiter from -50% to +25%, which was the case with RDNA2.

Suffice to say, I ended up buying a 4070S and it can run at 130-140W all day long with just ~10% performance sacrifice and barely any heat and noise. I ran some benchmarks, and the performance is basically on par with the 320W RTX3080.

I don't think it's a waste of silicon, no. If anything, it's their lifeline - they need this thing to succeed and they need this thing to sell. If they successfully replicate the things that made Polaris great - namely acceptable performance at a decent efficiency range and at a TDP tier that will fit on most budget systems without difficulty, they'll have a solid product in their hands. They clearly lack the clout to make a competent high-end product, so I think it's a good decision.

What must absolutely not happen is the continuous stability and BSOD problems that have plagued Radeon, it's really gotten bad to the point where people have begun to just consider it a normal thing and even factor it into the purchasing decision as a "meh, it was cheap and I don't mind if it crashes a little"... it's frustrating to no end.

Anyway, blocking tools like MPT's down to taking away control from the users, which are often irresponsible and go after third party snake oil (like modded drivers and VBIOS) without thinking of the consequences. That way they can lower the RMA rate and make their AIB partners happier. Signed VBIOSes, frequency and voltage range lockouts, hardware shunts - basically have become anti-counterfeiting and anti-customization security features. Nvidia does it too.

 

[AusWolf]

A waste of silicon...

RDNA4 would be just another trash tech unless they open up the firmware and allow the end user to tweak the cards to their liking.

As bad as RDNA3 was, the complete lack of MorePowerTool support was the last nail in its coffin... at least for me personally.

I briefly considered a 7800XT but was absolutely appalled by its undervolting behavior. While you can apply an undervolt, the card would just want to hit its power limiter and will use the extra headroom to squeeze out a few megahertz worth of frequency - which would be next to worthless - instead of saving power.

After all, the V/F curve doesn't scale linearly.

And to add insult to the injury, the power limiter on most 7800XTs only goes down by like 8% or so! On a 250W+ card! At least allow us to set the power limiter from -50% to +25%, which was the case with RDNA2.

Suffice to say, I ended up buying a 4070S and it can run at 130-140W all day long with just ~10% performance sacrifice and barely any heat and noise. I ran some benchmarks, and the performance is basically on par with the 320W RTX3080.

What's with some people getting so frantic about undervolting? What makes you think that you can make a better V/F curve than AMD's engineers? If I want a 50% less hungry card, I buy one that's a tier lower. That doesn't only save on my energy use, but also on the initial costs.

I did have a 7800 XT for a couple of months (I sold it because of financial priorities), and as a plug'n'play card, it was absolutely fine.

 

[Bad]

What must absolutely not happen is the continuous stability and BSOD problems that have plagued Radeon, it's really gotten bad to the point where people have begun to just consider it a normal thing and even factor it into the purchasing decision as a "meh, it was cheap and I don't mind if it crashes a little"... it's frustrating to no end.

Okay, first of all I've owned GPUs from both AMD and Nvidia. In fact, I currently have 3 GPUs from both and let me tell you; AMD's driver issues are way overblown.

People who complain about bad drivers either don't own their product, don't know how to properly install drivers (DDU is key), or are living in a fantasy world where AMD drivers are as bad as they used to be back in TeraScale days.

Anyway, blocking tools like MPT's down to taking away control from the users, which are often irresponsible and go after third party snake oil (like modded drivers and VBIOS) without thinking of the consequences.

I doubt that's such a widespread issue. Most people who dwell into BIOS modding usually know what they're doing.

I spent countless hours bothering people on various forums before attempting to flash my first custom vBIOS because I realized how risky that endavor was.

That way they can lower the RMA rate and make their AIB partners happier. Signed VBIOSes, frequency and voltage range lockouts, hardware shunts - basically have become anti-counterfeiting and anti-customization security features. Nvidia does it too.

I doubt any AIB partner is foolish enough to offer full refund on a bricked card with a custom vBIOS.

What's with some people getting so frantic about undervolting? What makes you think that you can make a better V/F curve than AMD's engineers?

I don't think you quite understand silicon binning.

Just because a certain card runs at X frequency at Y voltage doesn't mean all cards can hit that V/F curve.

If I want a 50% less hungry card, I buy one that's a tier lower. That doesn't only save on my energy use, but also on the initial costs.

So, instead of a 4070S I should've gone with the 4060Ti, a card that's 30% slower than the 4070S and has 4GB less vRAM to boot?!

Despite the aggressive underclock/undervolt, my 4070S is still easily ~20-25% faster than the 4060Ti a.k.a 3060Ti.

Plus, it's only drawing about 130-140W of power whereas the 4060 'Tie' is a 160W part.

Can't speak for everyone but I think (believe) the 4070S was well worth the ~$200 price premium. And don't get me started on the ~$500 4060Ti!

 

[AnotherReader]

What's with some people getting so frantic about undervolting? What makes you think that you can make a better V/F curve than AMD's engineers? If I want a 50% less hungry card, I buy one that's a tier lower. That doesn't only save on my energy use, but also on the initial costs.

I did have a 7800 XT for a couple of months (I sold it because of financial priorities), and as a plug'n'play card, it was absolutely fine.

I believe it's more a matter of having the ability to adjust the power limit. Older AMD GPUs had a range of 50% to 150% out of the box. For folding at home, the most efficient point is typically 50 to 65% of full power draw. I can set that easily with my Vega 64, but RDNA 3 has neutered this ability.

 

[AusWolf]

I believe it's more a matter of having the ability to adjust the power limit. Older AMD GPUs had a range of 50% to 150% out of the box. For folding at home, the most efficient point is typically 50 to 65% of full power draw. I can set that easily with my Vega 64, but RDNA 3 has neutered this ability.

F@H doesn't even use the full power of the card out of the box, even with a 100% limit, does it?

 

[AnotherReader]

F@H doesn't even use the full power of the card out of the box, even with a 100% limit, does it?

It depends upon the work unit. Some of the larger work units will get very close to 100% power consumption.

 

[DaemonForce]

What's with some people getting so frantic about undervolting? What makes you think that you can make a better V/F curve than AMD's engineers?

If you have to ask, you're not going to understand the answers.
You're not the audience of users that do upscaled 2K+AA 90FPS x2 and then experiencing random geometry errors, particle overload and texture crashes on random polys.
Can you even recognize the nightmare fuel?

The little firebug cheeseball doesn't normally look like this but gets there on stock settings with my RX 580.
This is why I tune it down to an extremely specific window of 1105mV.

 

[AusWolf]

If you have to ask, you're not going to understand the answers.
You're not the audience of users that do upscaled 2K+AA 90FPS x2 and then experiencing random geometry errors, particle overload and texture crashes on random polys.
Can you even recognize the nightmare fuel?

View attachment 370287

The little firebug cheeseball doesn't normally look like this but gets there on stock settings with my RX 580.
This is why I tune it down to an extremely specific window of 1105mV.

You shouldn't see graphical errors with any GPU running at stock. If you do, that only means your card is faulty.

 

[DaemonForce]

Oh yeah, faulty cherry picked enthusiast card that only experiences this problem in VR after reaching 65℃.
Or continuously crashes routine Firefox windows to desktop on junk drivers at stock voltages ramping up the temps to 65℃.
Absolutely zero connection there whatsoever. We're completely baffled!

 

[AusWolf]

Oh yeah, faulty cherry picked enthusiast card that only experiences this problem in VR after reaching 65℃.
Or continuously crashes routine Firefox windows to desktop on junk drivers at stock voltages ramping up the temps to 65℃.
Absolutely zero connection there whatsoever. We're completely baffled!

Do you see it anywhere in your card's manual, or the driver's release notes that you should use your card undervolted if you experience errors? Do you expect every single RX 580 owner to know how to do it? I'm the one that's baffled.

Yes, it is a faulty, cherry picked card, or one with bad drivers as far as I'm concerned.

 

[Marcus L]

Oh yeah, faulty cherry picked enthusiast card that only experiences this problem in VR after reaching 65℃.
Or continuously crashes routine Firefox windows to desktop on junk drivers at stock voltages ramping up the temps to 65℃.
Absolutely zero connection there whatsoever. We're completely baffled!

That card is old as dirt, had many back in the day and bios modded them, flashed them, undervolted them until the cows came home, they were a legendary card IMO back in the day but you are now comparing apples to bananas in this conversation about GPU's 8 years later, and FWIW I think you need to retire that, vRAM likely starting to degrade ofc you will slate me for not knowing what I am talking about and how good your 580 is and it's not on it's last legs etc etc I get it

Oh yeah, faulty cherry picked enthusiast card that only experiences this problem in VR after reaching 65℃.
Or continuously crashes routine Firefox windows to desktop on junk drivers at stock voltages ramping up the temps to 65℃.
Absolutely zero connection there whatsoever. We're completely baffled!

Literally answered your own questions, only does it when it heats up and is stressed and you are using EOL drivers

 

[DaemonForce]

Do you see it anywhere in your card's manual, or the driver's release notes that you should use your card undervolted if you experience errors?

No. I also don't see any other RX 580 users with a VR experience as good either. It's a factory overclocked card.
Why would PowerColor tell anyone to undervolt a Red Devil/OC?

Do you expect every single RX 580 owner to know how to do it?

You really want to pull that thread?
When half the spam on these boards is dumbdumb and dumbass going for the Chinese chopping block edition of this card with immediate regret? Fffhahahaha!
"Ermagherd, I paid liek $80 for this mystery Monday miner card that is old as balls and only the HDMI port works, send halp!"
Yeah, they won the prize.

That card is old as dirt, had many back in the day and bios modded them, flashed them, undervolted them until the cows came home, they were a legendary card IMO back in the day

QED. We all did it. We knew something wasn't right and how to experiment with this and that, even if solutions weren't one size fits all.

but you are now comparing apples to bananas in this conversation about GPU's 8 years later, and FWIW I think you need to retire that, vRAM likely starting to degrade ofc you will slate me for not knowing what I am talking about and how good your 580 is and it's not on it's last legs etc etc I get it

No I get it. It had one job and needed tuning just to behave but it also outlived that purpose without any real successor.
It had the best pairing with my FX-8370 and AMD 970 board, blocks on both parts. I should have kept them together but it is what it is.
Never water cooling the 580 because that's dumb and I'm now spoiled for "high end" choices, even if I have to wait another two months for option 3.
Whatever performance comes of the 8600XT or 880XT is going to get examined with several microscopes and I'm definitely going to be there for it Day 1.
Maybe I'll pick an 8000 series card, it could happen. It's also far more likely given my history and really nasty luck that has me fall back to this old card.
So yeah, I'm excited.

Literally answered your own questions, only does it when it heats up and is stressed and you are using EOL drivers

In 2018, every other driver did this and it was still a growing pain, similar to the RX 7000 series drivers leading up to now.
What I'm saying is they weren't EOL back then. They were bleeding edge and yes, some of us did bleed for it.
Today's drivers are much better but things are much different in this era and there's no way I can expect performance.
The biggest hurdle with games now is feature lockout. We've had years to get everyone on the same page with DX12_2 and there's no excuses anymore.

 

[Kapone33]

I don't think it's a waste of silicon, no. If anything, it's their lifeline - they need this thing to succeed and they need this thing to sell. If they successfully replicate the things that made Polaris great - namely acceptable performance at a decent efficiency range and at a TDP tier that will fit on most budget systems without difficulty, they'll have a solid product in their hands. They clearly lack the clout to make a competent high-end product, so I think it's a good decision.

What must absolutely not happen is the continuous stability and BSOD problems that have plagued Radeon, it's really gotten bad to the point where people have begun to just consider it a normal thing and even factor it into the purchasing decision as a "meh, it was cheap and I don't mind if it crashes a little"... it's frustrating to no end.

Anyway, blocking tools like MPT's down to taking away control from the users, which are often irresponsible and go after third party snake oil (like modded drivers and VBIOS) without thinking of the consequences. That way they can lower the RMA rate and make their AIB partners happier. Signed VBIOSes, frequency and voltage range lockouts, hardware shunts - basically have become anti-counterfeiting and anti-customization security features. Nvidia does it too.

What BSOD are you talking about? Do you have evidence of such? Are you claiming that no Nvidia card has suffered from that? What are you talking about taking away from users. You seem to have no idea what you are talking about. I wonder why you always bring a negative conjecture when you don't even have an AMD card to justify your comments. If you bring the 5000 series up again I will laugh. Not because it is funny but because I have a 5600XT that works just fine.

 

[AusWolf]

No. I also don't see any other RX 580 users with a VR experience as good either. It's a factory overclocked card.
Why would PowerColor tell anyone to undervolt a Red Devil/OC?

You apparently think that undervolting is a standard everyday procedure that must be done on an RX 580 to run properly. If that's the case then there must be some mention of it in the user's manual. Oh there isn't? I wonder why.

You really want to pull that thread?
When half the spam on these boards is dumbdumb and dumbass going for the Chinese chopping block edition of this card with immediate regret? Fffhahahaha!
"Ermagherd, I paid liek $80 for this mystery Monday miner card that is old as balls and only the HDMI port works, send halp!"
Yeah, they won the prize.

You didn't answer the question.

 

[Vayra86]

I don't think it's a waste of silicon, no. If anything, it's their lifeline - they need this thing to succeed and they need this thing to sell. If they successfully replicate the things that made Polaris great - namely acceptable performance at a decent efficiency range and at a TDP tier that will fit on most budget systems without difficulty, they'll have a solid product in their hands. They clearly lack the clout to make a competent high-end product, so I think it's a good decision.

What must absolutely not happen is the continuous stability and BSOD problems that have plagued Radeon, it's really gotten bad to the point where people have begun to just consider it a normal thing and even factor it into the purchasing decision as a "meh, it was cheap and I don't mind if it crashes a little"... it's frustrating to no end.

Anyway, blocking tools like MPT's down to taking away control from the users, which are often irresponsible and go after third party snake oil (like modded drivers and VBIOS) without thinking of the consequences. That way they can lower the RMA rate and make their AIB partners happier. Signed VBIOSes, frequency and voltage range lockouts, hardware shunts - basically have become anti-counterfeiting and anti-customization security features. Nvidia does it too.

Dafuq are you talking about, BSOD problems? I haven't had any. It also doesn't crash 'a little'. What Radeons will do is allow you to use the hardware as you see fit, and yes if you are a fool, you can indeed break it. But that's a PEBCAK problem. Not a Radeon problem, and it certainly doesn't come into play if you just plug and play the thing.

Literally everything I've thrown at my 7900XT, it ate for breakfast.

Nvidia is no different in how this all works, if you're not stable, it will simply blackscreen on you much earlier, not showing the anomalies on screen, but reverting to stable clocks/volts.

If you have to ask, you're not going to understand the answers.
You're not the audience of users that do upscaled 2K+AA 90FPS x2 and then experiencing random geometry errors, particle overload and texture crashes on random polys.
Can you even recognize the nightmare fuel?

View attachment 370287

The little firebug cheeseball doesn't normally look like this but gets there on stock settings with my RX 580.
This is why I tune it down to an extremely specific window of 1105mV.

If you have this issue while undervolting, you're not running a stable undervolt.

The End.

I saw this with my 7900XT too, I had it running at 1025mv for months and had the odd crash while it would run everything for hours just fine. Then I ran Black Myth Wukong on it and it insta-crashed every time on that voltage; upped it to 1075mv and now it runs everything flawlessly, no more odd crashing either, and no intermittent graphical anomalies either.