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.

 

[Glo.]

Wasn't TSMC rumored to be "forcing" everybody to 6 nm process since they will have more wafers available to everybody, because of the die shrink that it allows, and supposedly is cheaper than N7?

 

[Ajay]

Wasn't TSMC rumored to be "forcing" everybody to 6 nm process since they will have more wafers available to everybody, because of the die shrink that it allows, and supposedly is cheaper than N7?

I think the word was 'encouraging'. It's not a zero dollar switch. Re-validation and new masks at the minimum.

 

[Hitman928]

Wasn't TSMC rumored to be "forcing" everybody to 6 nm process since they will have more wafers available to everybody, because of the die shrink that it allows, and supposedly is cheaper than N7?

They are pushing (but not forcing) all new designs to 6 nm. They are still running the 7 nm lines and those designs that have already taped out on 7 nm will keep going.

 

[blckgrffn]

If I were AMD I’d be tempted to just move the current lineup naming scheme to XTX on 6nm parts, raise clock speeds very superficially and just take smaller die sizes for more margin and pass along a mild power savings as well. If possible for board partners have them reuse the same PCB and everything.

They all sell out regardless.

 

[Mopetar]

It's all a matter of economics. If you don't plan on keeping a particular product around for too much longer then moving it to 6nm doesn't make financial sense as the density savings aren't enough to make up for the initial investment. If you have something that's using a lot of wafers or is going to be manufactured for years to come then a move makes sense.

RDNA2 is a bit more interesting since SRAM doesn't shrink as well as logic and the more Infinity Cache a chip has, the lower the gains are. I dint think any of them have so much cache that this would be a deal breaker, but it does factor in.

 

[Glo.]

N6 as far as I know requires less masks and less steps in the manufacturing process resulting in better cost efficiency. You have higher wafer output and are getting more dies per wafer.

 

[VirtualLarry]

Would they be refreshing the high-end RDNA2 dies (more performance), or the low-end (for cost savings, more dies per wafer because higher density), or both? I wonder where the true economies of scale lie.

It would be kind of weird, maybe, if Navi 24 was made on TSMC N6, and all the others continued on 7. That would be more like a sort of risk-production / testing the waters sort of thing, woudn't it?

Kind of like Bonaire, years before.

 

[Frenetic Pony]

Probably mid to low end right? High end has plenty of power headroom, the 6900xt just needs faster ram (which I guess Samsung can provide).

Besides, mid to low end also fits in laptops. Why do high end when RDNA3 will be bringing much better high end desktop stuff to market like 6 months later, but maybe a year later for laptop parts. But I guess it matters how much it costs to switch and how much volume is available on each.

 

[jpiniero]

Probably mid to low end right? High end has plenty of power headroom, the 6900xt just needs faster ram (which I guess Samsung can provide).

Besides, mid to low end also fits in laptops. Why do high end when RDNA3 will be bringing much better high end desktop stuff to market like 6 months later, but maybe a year later for laptop parts. But I guess it matters how much it costs to switch and how much volume is available on each.

I'd hate to see what the prices are going to be for the chiplet products.

 

[NostaSeronx]

6nm EUV(Re-Tape-Out):
Lower lead times do to single-pattern EUV mask; >275 WPH / 4 = >68.75 WPH versus >155->170 WPH for EUV machine

More accurate to simulation printing from single-pattern EUV; Significantly less shorts on wires, etc.

Porting the design to 6nm is a quick way to get a more reliable yielding supply at lower lead times. There is no reason not to expect a complete refresh Zen2/Zen3/RDNA1/CDNA1/XSX/XSS/PS5 going to 6nm as well.

6nm EUV(New-Tape-Out): 7nm Plus like density, primarily for new products not refreshes.

 

[beginner99]

Why do high end when RDNA3 will be bringing much better high end desktop stuff to market like 6 months later,

Because you can use them as "upper mid-range" in the next series. Eg 7800xt/7900xt and maybe 7700xt can be new 5nm products while the 6800xt gets "rebranded" to 7600xt. It's not really a rebrand as it's the same chip on 6nm but you get the point. 5nm capacity will be limited. no way anything else but high-end will be on it at start and also no need to steal more wafers from zen4 server cpus.

 

[soresu]

Kind of like Bonaire, years before.

Bonaire was a different process node?

I thought it was just a different µArch (Sea Islands/GCN2) vs the rest of the HD 7xxx series (Southern Islands/GCN1).

 

[soresu]

There is no reason not to expect a complete refresh Zen2/Zen3/RDNA1/CDNA1/XSX/XSS/PS5 going to 6nm as well.

Why Zen2, RDNA1 or CDNA1?

More likely they will just EOL those products and release only Zen 3, RDNA2 and CDNA2 products going forward.

I suspect that the only Zen2 designs getting 6nm ports will be the console parts including Van Gogh for the Steam Deck.

 

[NostaSeronx]

Why Zen2, RDNA1 or CDNA1?

More likely they will just EOL those products and release only Zen 3, RDNA2 and CDNA2 products going forward.

I suspect that the only Zen2 designs getting 6nm ports will be the console parts including Van Gogh for the Steam Deck.

Reduction of cost from >88 masks to >66 masks. This gives cost savings to existing contracts using V2000, MI100, BC-100, etc series lines.

6nm EUV approaches the same price as 16nm/14nm with 6nm EUV increasing wspm for TSMC

1. TSMC reducing the price of 6nm to 10nm through 12nm wafer costs. = $4000-$5000 versus $6000-$7000.
2. Given 100,000 wsp arb time from 7nm to 150,000 wsp arb time for 6nm. :: 275 wph to >55 wph(SAQP) and >92 wph(SADP) to >170 wph(EUV SP(good till 28nm-Mx))

 

[Olikan]

6nm EUV(Re-Tape-Out):
Lower lead times do to single-pattern EUV mask; >275 WPH / 4 = >68.75 WPH versus >155->170 WPH for EUV machine

More accurate to simulation printing from single-pattern EUV; Significantly less shorts on wires, etc.

Porting the design to 6nm is a quick way to get a more reliable yielding supply at lower lead times. There is no reason not to expect a complete refresh Zen2/Zen3/RDNA1/CDNA1/XSX/XSS/PS5 going to 6nm as well.

6nm EUV(New-Tape-Out): 7nm Plus like density, primarily for new products not refreshes.
View attachment 54352

There was any talk about EUV improving Fmax?

 

[NostaSeronx]

There was any talk about EUV improving Fmax?

EUV improves Fmax by improving wire characteristics. SAQP has a very small margin of error from GDS (bumpy). Whereas EUV has fold smaller margin of error from GDS (unbumpy).

Example for Fmax area:


RDNA2 and Zen2/Zen3 are heavily Fmax-orientated already. So, getting them on 6nm which prints M1/M2/M3 better makes sense.

In custom designs the N6 transistor can have the N5 Perf-booster as well. Which can mean 6nm products can hit lower TDPs as well or higher clocks at same TDP.

Up to 1.1x Freq(some perf lost to binning for most yielded dies);
CPU-side -- Cezanne-59-"80/00" 4.8/4.6 GHz -> Rembrandt-69-"80/00" 5.0/4.9 GHz
GPU-side-example -- 6-"700/800"M 2.3 GHz -> 6-"755/85"M <2.5 GHz.
~0.8x Power => have to wait for low TDP variants.
Ex: VGH-FF3 7nm (4W) -> MER-FF3 6nm (>3.2W)

Edit: for certain products still being 7nm.

 

[RnR_au]

Seems upgrades are indeed coming... Q2 apparently...

https://twitter.com/x/status/1483374668719738880

 

[Shmee]

Huh, sounds potentially promising, hopefully it is true, especially if the performance is really good, and a different process such as 6nm can be used, which could allow for more product production, and therefore better availability and pricing.

 

[Insert_Nickname]

and a different process such as 6nm can be used, which could allow for more product production, and therefore better availability and pricing.

6nm might allow AMD to squeeze out a few more chips per wafer, so that'll help a bit. I doubt it'll have much effect on pricing with how things are.

 

[scineram]

This is not a thing. Just saying.

 

[Mopetar]

6nm might allow AMD to squeeze out a few more chips per wafer, so that'll help a bit. I doubt it'll have much effect on pricing with how things are.

Isn't the TSMC 6nm process done using wafers that would otherwise have used 7nm? If that's the case, absent of any additional production capacity that AMD managed to acquire, I'm not sure that this helps with availability. Really it would just mean better performance characteristics, which while nice, don't help much if you can't get a GPU in the first place.

This is not a thing. Just saying.

I get that it's a rumor and those aren't worth a lot, but do you have anything substantive to add to this?

 

[Hitman928]

Isn't the TSMC 6nm process done using wafers that would otherwise have used 7nm? If that's the case, absent of any additional production capacity that AMD managed to acquire, I'm not sure that this helps with availability. Really it would just mean better performance characteristics, which while nice, don't help much if you can't get a GPU in the first place.



I get that it's a rumor and those aren't worth a lot, but do you have anything substantive to add to this?

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.

 

[scineram]

Why would AMD tapeout new chips at great expense for such minimal gain?
And even if they did where are the codenames? The rumormill has nothing afaik.

 

[RnR_au]

Isn't the TSMC 6nm process done using wafers that would otherwise have used 7nm? If that's the case, absent of any additional production capacity that AMD managed to acquire, I'm not sure that this helps with availability.

N6 manufacturing lines have more wafer/month capacity than N7 manufacturing lines. Together with the higher density on the wafers it may make an impact at retail, if all the other card components are readily available at a fair price. Also note that TSMC have more N6 capacity than N7 capacity now, so would be encouraging their larger customers to shift across.

But yeah... as scineram said above, would they do it? The rumour says Q2... does that mean RDNA3 is 2023, or is AMD being aggressive for market share?