Discussion RDNA 2 6nm Refresh

[GodisanAtheist]

AMD Allegedly Preparing Refreshed 6 nm RDNA 2 Radeon RX 6000S GPU

AMD is allegedly preparing to announce the Radeon RX 6000S mobile graphics card based on a refreshed RDNA 2 architecture. The new card will be manufactured on TSMC's N6 process which offers an 18% logic density improvement over the N7 process currently used for RDNA 2 products resulting in...

www.techpowerup.com

- 6nm would bring 18% higher density (so more dies per wafer) + some performance or efficiency improvements as AMD chooses to tune the dies (Performance for desktop/Efficiency for laptops I suppose).

My understanding is that 6nm is a clean shrink of 7nm without the complexities and overhead of a full node shrink.

Wonder if AMD will end up running a couple simultaneous lines: 7nm for their 6000 series mainstream and bulk cards, 6nm for their mid to top end 6000 series S cards, and eventually 5nm for their RDNA 3 top end parts. Shift everything down a tier when 4/3nm whatever show up.

 

[Mopetar]

6nm has density improvements which could increase chips per wafer, assuming AMD actually used it to improve density. The other thing is that 6 nm has reduced mask layers which should allow for more wafers to be produced in the same time frame as 7 nm which of course could lead to more chips per month coming out of the fab for AMD.

Would they necessarily make a new set of masks though? Even if the process requires fewer masks, it would still be expensive. The expectation is that RDNA 3 GPUs will be releasing towards the end of the year or early next so that also calls into question why AMD would produce a new set of masks. Maybe it is worth it if they can ramp up production enough, but they can also just take advantage of the node to offer better clock speeds or improved efficiency even without taking advantage of a density increase. They've been quite close to Nvidia in terms of performance across their lineup so a small bump in performance is going to help them just as much as anything else.

 

[Hitman928]

Would they necessarily make a new set of masks though? Even if the process requires fewer masks, it would still be expensive. The expectation is that RDNA 3 GPUs will be releasing towards the end of the year or early next so that also calls into question why AMD would produce a new set of masks. Maybe it is worth it if they can ramp up production enough, but they can also just take advantage of the node to offer better clock speeds or improved efficiency even without taking advantage of a density increase. They've been quite close to Nvidia in terms of performance across their lineup so a small bump in performance is going to help them just as much as anything else.

I don't know from a business sense if it makes sense for a 6 nm refresh at this point. It could potentially if they are doing it with limited SKUs to keep as the lower end after RDNA3 comes out. If the upper metal routing doesn't need to change and assuming they are compatible with N6 and AMD doesn't care about increasing density, then they wouldn't have to buy those masks again.

 

[Mopetar]

I don't know from a business sense if it makes sense for a 6 nm refresh at this point. It could potentially if they are doing it with limited SKUs to keep as the lower end after RDNA3 comes out. If the upper metal routing doesn't need to change and assuming they are compatible with N6 and AMD doesn't care about increasing density, then they wouldn't have to buy those masks again.

From prior posts about N6, my understanding was that it's compatible with N7 masks so there's no need to get a new set made. They really don't lose anything moving to N6 outside of the ability to use it for anything else that they could make on it. Any clock speed bump they can get is a nice bonus.

AMD probably would give them a different name in that case, but I don't think people with RDNA 2 cards would upgrade outside of people who just want the best new AMD card. Frankly the supply for the original cards from AMD was bad enough that there probably aren't a lot of people who could upgrade anyhow.

I suppose that's a refresh.

 

[beginner99]

I don't know from a business sense if it makes sense for a 6 nm refresh at this point

they probably don't know either or maybe they actually do but it's clear with the release schedule to have something new as well when intel will launch arc.

 

[jpiniero]

I don't know from a business sense if it makes sense for a 6 nm refresh at this point. It could potentially if they are doing it with limited SKUs to keep as the lower end after RDNA3 comes out. If the upper metal routing doesn't need to change and assuming they are compatible with N6 and AMD doesn't care about increasing density, then they wouldn't have to buy those masks again.

Aren't the rumors saying that RDNA3 chiplet models are going to be high end only, at least initially? And who knows what prices constitute "High End". A medium effort refresh would be helpful in filling out the rest of the stack.

 

[Hitman928]

Aren't the rumors saying that RDNA3 chiplet models are going to be high end only, at least initially? And who knows what prices constitute "High End". A medium effort refresh would be helpful in filling out the rest of the stack.

I don't know, I don't follow the rumors too much anymore, but yes, RDNA3 mid/high end with RDNA2 refresh mid/low end might make sense.

 

[Mopetar]

Almost every GPU launch starts out with the high end products first regardless of what comes after. With shipping and other component costs being a lot higher, it makes even more sense to have the new take focus on high end parts where those increases are a smaller overall percentage and more easily absorbed.

All the same I'd still probably refresh the whole RDNA2 lineup if I were AMD regardless of what the plans are for RDNA3 GPUs that are coming out this year.

 

[Hitman928]

Almost every GPU launch starts out with the high end products first regardless of what comes after. With shipping and other component costs being a lot higher, it makes even more sense to have the new take focus on high end parts where those increases are a smaller overall percentage and more easily absorbed.

All the same I'd still probably refresh the whole RDNA2 lineup if I were AMD regardless of what the plans are for RDNA3 GPUs that are coming out this year.

With the crazy amount of GPU demand and prices, I don't see why AMD can't have both RDNA2 refresh and RDNA3 running at the same time and for a while yet. They wouldn't even have to lower RDNA2 prices since street price is so much higher than MSRP anyway. Maybe they'll have to not include the 6900XT in the refresh. They could then set RDNA3 at higher MSRP. Since they are on different nodes, you could produce a lot more GPUs in total and grow margin at the same time with higher RDNA3 prices. Keep this up until things finally start cooling down and supply meets demand at which point you phase out RDNA2 or just lower prices across the board. We'll see what AMD's actual plan is.

 

[Ajay]

With the crazy amount of GPU demand and prices, I don't see why AMD can't have both RDNA2 refresh and RDNA3 running at the same time and for a while yet. They wouldn't even have to lower RDNA2 prices since street price is so much higher than MSRP anyway. Maybe they'll have to not include the 6900XT in the refresh. They could then set RDNA3 at higher MSRP. Since they are on different nodes, you could produce a lot more GPUs in total and grow margin at the same time with higher RDNA3 prices. Keep this up until things finally start cooling down and supply meets demand at which point you phase out RDNA2 or just lower prices across the board. We'll see what AMD's actual plan is.

I think this would be a good plan, though keeping RDNA2 on 7nm is likely most cost effective for such a purpose. However, we are at the mercy of AMD (and other GPU makers) as to what they actually do.

 

[Hitman928]

I think this would be a good plan, though keeping RDNA2 on 7nm is likely most cost effective for such a purpose. However, we are at the mercy of AMD (and other GPU makers) as to what they actually do.

If they actually plan on keeping RDNA2 cards around a while, moving to 6 nm makes sense with the given volume increase it provides. The (partial) mask set cost will be trivial if the refresh stays in production long enough to live alongside RDNA3.

 

[GodisanAtheist]

There are rumors abound of an AMD "xx50" line refresh of all of their 7nm parts down to 6nm.

By all accounts 6nm does not require any significant rework to get going, I kinda feel like it would be a no brainer for AMD.

I anticipate RDNA3 and Lovelace are both getting pushed back or getting a very gradual ramp. Until supply issues ease up, releasing a new series just puts a crimp on sales from the prior gen while still being unable to feed the current market.

 

[Frenetic Pony]

There are rumors abound of an AMD "xx50" line refresh of all of their 7nm parts down to 6nm.

By all accounts 6nm does not require any significant rework to get going, I kinda feel like it would be a no brainer for AMD.

I anticipate RDNA3 and Lovelace are both getting pushed back or getting a very gradual ramp. Until supply issues ease up, releasing a new series just puts a crimp on sales from the prior gen while still being unable to feed the current market.

Nah, this doesn't mean RDNA3 is being delayed. It'll probably be out, Novemberish? And until then why not charge the same amount of money you're already getting, or more, for products with somewhat better yields? Higher profit margins all around for three quarters probably sounds worth the engineering cost of a minor refresh.

 

[beginner99]

With the crazy amount of GPU demand and prices, I don't see why AMD can't have both RDNA2 refresh and RDNA3 running at the same time and for a while yet. They wouldn't even have to lower RDNA2 prices since street price is so much higher than MSRP anyway. Maybe they'll have to not include the 6900XT in the refresh. They could then set RDNA3 at higher MSRP. Since they are on different nodes, you could produce a lot more GPUs in total and grow margin at the same time with higher RDNA3 prices. Keep this up until things finally start cooling down and supply meets demand at which point you phase out RDNA2 or just lower prices across the board. We'll see what AMD's actual plan is.

I'm not sure it helps much with volume. AMD could make much more already now if they wanted to at cost of CPUs. Since CPUs are much more profitable even with current CPU pricing....the real relive would be Zen4 on 5nm so that most 6/7nm is free for GPU.

I think this would be a good plan, though keeping RDNA2 on 7nm is likely most cost effective for such a purpose. However, we are at the mercy of AMD (and other GPU makers) as to what they actually do.

All rumors hint to 6nm as supposedly TSMC has also been incentivizing customers to move. Since it takes less layers it should be much faster to manufacture and hence increases throughput. and with small density improvement you also can some extra dies per wafer. Pulling this number put of my ass but I would expect around 30% increased production rate if everyone moved from 6 to 7nm. This would help a lot. and of course more wafer = more profits for TSMC

 

[Hitman928]

I'm not sure it helps much with volume. AMD could make much more already now if they wanted to at cost of CPUs. Since CPUs are much more profitable even with current CPU pricing....the real relive would be Zen4 on 5nm so that most 6/7nm is free for GPU.

I was thinking RDNA3 would be on 5 nm but after review, AMD just said it would be on an advanced node so if it stays on 6 nm then yeah, it won't help with volume.

 

[Ajay]

I think long term volume production of RDNA2 isn't something AMD want, so the costs of even a small redesign and re-verification probably wouldn’t be worth it; despite what I or the community might like, IMHO. Heck, I’d love a reasonably priced 6800! [even though I'm a nvidiot]

 

[eek2121]

Amazon has 6900XTs in stock (shipped from sold by, still not at the MSRP, they are $1,500, but that is a decent price IMO, considering that is about what I paid for my 3090). I almost pulled the trigger (came REALLY close). but knowing we COULD see a refresh has stopped me from doing it.

 

[Ajay]

Amazon has 6900XTs in stock (shipped from sold by, still not at the MSRP, they are $1,500, but that is a decent price IMO, considering that is about what I paid for my 3090). I almost pulled the trigger (came REALLY close). but knowing we COULD see a refresh has stopped me from doing it.

$1500 - $1800 used to be my full new computer budget including a good GPU. I built my current system for a bit less than that, but with an older GPU. So, pardon me for thinking things are pretty insane right now.

 

[jpiniero]

Amazon has 6900XTs in stock (shipped from sold by, still not at the MSRP, they are $1,500, but that is a decent price IMO, considering that is about what I paid for my 3090). I almost pulled the trigger (came REALLY close). but knowing we COULD see a refresh has stopped me from doing it.

The MSRP of the 6900 XT is $999. Like say they up the clocks 10% and use 18 gbps memory. Surely they could make the MSRP whatever they want, and even after RDNA3 launches they could maybe "cut it" to $999.

 

[GodisanAtheist]

The MSRP of the 6900 XT is $999. Like say they up the clocks 10% and use 18 gbps memory. Surely they could make the MSRP whatever they want, and even after RDNA3 launches they could maybe "cut it" to $999.

- That's the gameplan here. Push MSRPs as high as possible this gen, look like saints and have all the simps eating out of your hand the next when you "slash" MSRP back down to 25% above the prior gen's launch MSRP.

Basically EXACTLY what happened with NV between the 1xxx/2xxx/3xxx series. 3xxx series looked incredible largely due to 2xxx series having straight garbage price to performance.

 

[eek2121]

$1500 - $1800 used to be my full new computer budget including a good GPU. I built my current system for a bit less than that, but with an older GPU. So, pardon me for thinking things are pretty insane right now.

The MSRP of the 3090 is $1,499.99. The 6900 XT is as good as or better than the 3090. The only area it is really behind in is in ray tracing. IMO, it is a fair price for such a high end card. You can get cheaper cards, it just depends on your budget. You should get ready, because prices will never come back down to levels where they are supposed to be.

- That's the gameplan here. Push MSRPs as high as possible this gen, look like saints and have all the simps eating out of your hand the next when you "slash" MSRP back down to 25% above the prior gen's launch MSRP.

Basically EXACTLY what happened with NV between the 1xxx/2xxx/3xxx series. 3xxx series looked incredible largely due to 2xxx series having straight garbage price to performance.

This isn't some conspiracy. Supply chain and BOM issues are causing high prices. Yes, prices are never going to be that low again. Inflation is over 7% and even when supply chain issues get sorted out, companies have realized they are just throwing money away to scalpers. 5nm GPUs also will cost AMD and NVIDIA significantly more since die sizes aren't getting smaller.

 

[CakeMonster]

I could see AMD doing this if RDNA3 is late Q4 or Q1, but I wouldn't put money on it. Several prerequisites will need to be present. They need to actually save significant physical size of the chip (cost) which AFAIK is not a given because of the big cache that doesn't shrink that easily. The actual design job is not supposed to be hard in this case but what do I know. Also its a game with regards to competition with NV, they might or might not want to put out new chips that doesn't have improved RT or better DLSS like capabilities, which could make them look outdated.

 

[coercitiv]

You should get ready, because prices will never come back down to levels where they are supposed to be.

Reminds me of an old post from this forum essentially telling us to get used to the new cheap TIM in mainstream CPUs, solder is gone and never coming back. "Never" expired 1 year later.

 

[maddie]

Reminds me of an old post from this forum essentially telling us to get used to the new cheap TIM in mainstream CPUs, solder is gone and never coming back. "Never" expired 1 year later.

Completely misused analogy.

Think of the stories of when X used to cost $Y. We have gotten accustomed to a very long period of low prices. Monetary policy is driving this and there is no end in sight. Changing policy will cause a crash and prices to crater, but only until available stock depletes, as a lot of us won't have disposable $. The glory days of the $100 4850 happened with the crash of 2008. Not a coincidence.

 

[coercitiv]

Think of the stories of when X used to cost $Y. We have gotten accustomed to a very long period of low prices. Monetary policy is driving this and there is no end in sight. Changing policy will cause a crash and prices to crater, but only until available stock depletes, as a lot of us won't have disposable $. The glory days of the $100 4850 happened with the crash of 2008. Not a coincidence.

I grew up through periods of high inflation, in fact my "Hello world!" attempt on a personal computer was an "Inflation calculator", the situation was so dire that even little kids knew what the term meant. Since then I've also seen the effect of scalping in the international market on common goods such as food supplies. It's fascinating what you can do with the proper finances and insight into how the supply chain works. Inflation is real and we're not getting around it, but current GPU prices aren't influenced by inflation alone, and more importantly verdicts such as "prices will never come back down" are a very bad idea in a time when the market is subjected to immense forces outside of it's control, making room for scalping to trickle down the production chain. We don't care about prices in absolute form, we only care about how they relate to our purchasing power, and here there's a definite possibility that GPU prices will adjust dramatically for the better.

Don't know if anyone doubted Arachnotronic on that TIM prediction, but I certainly didn't (at least not enough to say otherwise). It was a market reality that we all had to get adjusted to, or so we thought. Except the prediction was made during a volatile state of the CPU market, and it aged badly. So we can agree to disagree on the future of GPU prices, but the analogy stands.

 

[maddie]

I grew up through periods of high inflation, in fact my "Hello world!" attempt on a personal computer was an "Inflation calculator", the situation was so dire that even little kids knew what the term meant. Since then I've also seen the effect of scalping in the international market on common goods such as food supplies. It's fascinating what you can do with the proper finances and insight into how the supply chain works. Inflation is real and we're not getting around it, but current GPU prices aren't influenced by inflation alone, and more importantly verdicts such as "prices will never come back down" are a very bad idea in a time when the market is subjected to immense forces outside of it's control, making room for scalping to trickle down the production chain. We don't care about prices in absolute form, we only care about how they relate to our purchasing power, and here there's a definite possibility that GPU prices will adjust dramatically for the better.

Don't know if anyone doubted Arachnotronic on that TIM prediction, but I certainly didn't (at least not enough to say otherwise). It was a market reality that we all had to get adjusted to, or so we thought. Except the prediction was made during a volatile state of the CPU market, and it aged badly. So we can agree to disagree on the future of GPU prices, but the analogy stands.

"We don't care about prices in absolute form, we only care about how they relate to our purchasing power, and here there's a definite possibility that GPU prices will adjust dramatically for the better."

I agree, but my thoughts are they won't return to yesterday's prices. if you do have experience with inflation, you might agree that wages generally lag prices.

For semiconductors, baring a breakthrough, which by definition is unforeseen, perf/$ costs seem to reached a low and are now rising due to increased transistor costs alone, even if you exclude the other factors in general inflation. Chiplets is a strategy to slow this trend.