AMD Radeon RDNA2 Navi (RX 6500, 6600, 6700, 6800, 6900 XT)

AMD/Intel's agreement has restrictions with respect to sub licensing which is why the console model changed to direct chip purchases with the PS4/Xbox One from a design licensing model.

But this is something that isn't very clear cut depending on how the agreements are written up which is the problem with this statement -

Since it's based on a false premise that AMD's wafer allocation that they secured or even TSMC capacity and ramp up would be the exact same if the consoles hypothetically didn't exist.
Also aren't the Navis on N7P, not N7? At least WikiChip claims N7P for Navi10. Surely N7 and N7P don't share the capacity
 
Also aren't the Navis on N7P, not N7? At least WikiChip claims N7P for Navi10. Surely N7 and N7P don't share the capacity
Can't say for sure about N7 and N7P but based off of older nodes, it most likely does share the same capacity.
Same raw wafers running through the same tools (for the most part) with just different recipes/techniques at certain points.
From what I've heard, N7P is just optimization through maturation.

Edit- Found an Anandtech article on 7NP-
https://www.anandtech.com/show/14687/tsmc-announces-performanceenhanced-7nm-5nm-process-technologies
and Wikichips-
https://fuse.wikichip.org/news/2567/tsmc-talks-7nm-5nm-yield-and-next-gen-5g-and-hpc-packaging/


And TSMC is supposed to be doing ~140-150k 7nm wafer starts a month.
 
Last edited:
And TSMC is supposed to be doing ~140-150k 7nm wafer starts a month.
Indeed, and in H1 2020 AMD was allocated 21% of TSMC capacity, about 30k wpm of N7.
In H2 AMD got some more wpms and in Q4 they took over Huawei allocation and currently around 40-45K wpm. Which is also what most reports out of TW indicate.
https://www.techpowerup.com/262590/amd-to-outpace-apple-as-tsmcs-biggest-7nm-customer-in-2020
45k wpm spread thin across all the product lines.
Even though Lisa did not give the numbers, she said it clearly in last Earnings calls (after the Xilinx announcement), they have more capacity in Q4 but supply is very tight. If not for Huawei relinquishing their contract (for whatever reasons) 30K wpm is what AMD would have been having and would have been even more dire than what it already is.

She did mention that they are working on it through 2021, so perhaps come Q1 21 things will improve, either due to added capacity which is corroborated in TSMC Earnings report of new investment in capacity increase. And also some of their other clients moving to N5 (Which is in a new fab and not sharing capacity with N7)
 
Last edited by a moderator:
I know there is not a lot of other solutions, but the ultra dependance on TSMC is kind of worrisome. They can't produce for everyone... Will it be even worse with 5nm ?
 
Will it be even worse with 5nm ?
It is more fun at N5 in 2021
Apple, NV, AMD, Xilinx, QCM, MTK and others fighting for 55-60K total wpm of N5(P) compared to ~140K wpm of N7 currently:LOL:. Granted most of the other folks outside of Mobile will migrate to N5 towards H2 2021. Still with the next Apple silicon also still using N5(P) in H2 21 ... things are bleak.
TSMC new fab doing N3 is coming online only in H2 2022.
From recent TSMC earnings report, they have set aside capital since couple of quarters for N5 capacity increase. but very unlikely they will hit even 80K wpm by end of 2021
 
What should be our question is why with so much market until TSMC is the one supplying the demand. Like Intel could enter and compete and so does SS. Of course it's not easy but if there is money there will be investment trying to get it.
 
What should be our question is why with so much market until TSMC is the one supplying the demand. Like Intel could enter and compete and so does SS. Of course it's not easy but if there is money there will be investment trying to get it.
Crazy thing right now is that even Intel is capacity limited, with its almost 400K wpm (300mm equivalent) of 14++ and 10+/SF, (from icinsights, no split given unfortunately)
I won't be surprised if NV sticks with Samsung Foundry and move up to 5LPE instead. Exynos 1080 is already in mass production for next flagship. They will have 1 year to improve yields, which, if you ignore TW media and read from Korean media, is not terrible as made out to be even in its current state.
Also Raja Koduri attending Samsung SAFE event could possibly hint about Xe having some parts of the chip fabbed at Samsung. But this is just my guess which is contrary to reports from TW which says bits of Xe will be fabbed at TSMC.
Also QCM is rumored to be moving some of their APs to 5LPE

Going all N5 is going to bottleneck AMD, not that they have a choice. A lot of their product stack would be behind at N7 with some halo SKUs only at N5.
Francois Piednoel famously calculated this all the way from 2018 or is it 2017?
RDNA3/CDNA2/Zen4 are all confirmed to be using N5P ("Advanced node " in AMD slides), and reports from TW says AMD has requested many customizations on top of N5P. Most of the source is RetiredEngineer on Twitter who translated the reports from TW (digitimes et al)

One consequence of this, especially on CPU side is that AMD allegedly would have to make compromise how big they can get their Zen core to be, compared to M1 for example, simply because of being constrained by silicon real estate.
 
Also aren't the Navis on N7P, not N7? At least WikiChip claims N7P for Navi10. Surely N7 and N7P don't share the capacity

Off hand I believe only TSMC 7nm which uses EUV (not sure if it's still branded as 7nm+?) for certain layers is separate (in large part due to the afore mentioned use of EUV). Otherwise the rest are iterative variants of the original 7nm DUV process.

I know there is not a lot of other solutions, but the ultra dependance on TSMC is kind of worrisome. They can't produce for everyone... Will it be even worse with 5nm ?

5nm is supposed to scale up faster than 7nm did likely in part due to the higher forecasted demand. In terms of AMD's product stack the rumors do suggest that they'll stay on 7nm for longer than just from 2021 through to 2022 at least for a sizable segment (if not majority) of their actual shipping products. But AMD transitioning it's product stack and the consumer situation has some differentiation.

In terms of the end consumer supply issue that's currently happening it's more complicated as there are quite a lot of factors (many of which are one offs, or likely are) contributing to both demand and supply shifts that weren't forecasted far back. The last cycle had similar supply/demand issues for consoles and would've put strain into supply chains. However things like Intel's situation, Samsung was likely expected to be further along as well (with respect to 7nm), Covid, other parts of the supply chain and etc. are also adding to the issues.

Also with regards to the sentiment of TSMC I think it's worth noting that their current situation may or may not just be cyclical. I think just roughly was it 6 years ago (? would have to look it up) but their were serious concerns that Apple would transition back to Samsung, and that Samsung would supplant TSMC has #2 behind Intel. Of course now the sentiment regarding all three is completely different.
 
Last edited:
Indeed, and in H1 2020 AMD was allocated 21% of TSMC capacity, about 30k wpm of N7.
In H2 AMD got some more wpms and in Q4 they took over Huawei allocation and currently around 40-45K wpm. Which is also what most reports out of TW indicate.
https://www.techpowerup.com/262590/amd-to-outpace-apple-as-tsmcs-biggest-7nm-customer-in-2020
45k wpm spread thin across all the product lines.
Even though Lisa did not give the numbers, she said it clearly in last Earnings calls (after the Xilinx announcement), they have more capacity in Q4 but supply is very tight. If not for Huawei relinquishing their contract (for whatever reasons) 30K wpm is what AMD would have been having and would have been even more dire than what it already is.

She did mention that they are working on it through 2021, so perhaps come Q1 21 things will improve, either due to added capacity which is corroborated in TSMC Earnings report of new investment in capacity increase. And also some of their other clients moving to N5 (Which is in a new fab and not sharing capacity with N7)

Eh, didn't AMD say they more than doubled their 7nm capacity from 1H'20? So it should be more like 60-80k wpm if the 30k is correct for 1H'20.
Also, they picked up capacity not just from Huawei losing allocation but Apple switching to 5nm.

We are supposed to be getting a 5nm TSMC fab here in Arizona. Construction starts in '21 and completed in '23. But slim pickings with only 20k wpm moving in '24, chumpchange.
 
Last edited:
Eh, didn't AMD say they more than doubled their 7nm capacity from 1H'20? So it should be more like 60-80k wpm if the 30k is correct for 1H'20.
Also, they picked up capacity not just from Huawei losing allocation but Apple switching to 5nm.

We are supposed to be getting a 5nm TSMC fab here in Arizona. Construction starts in '21 and completed in '23.
Around 45K wpm seems to be the consistent figure for Q4 20, unfortunately I cant read chinese and I relied on RetiredEngineer translation.

End of Q4 2019 AMD had 20K wpm (At this time AMD's products still has a lot off 14nm), in H1 2020, (around mid Q2 ), AMD became the biggest customer of N7 getting 30k wpm due Apple ramping up A14 production. ( which can be verified with TSMC Q2 earnings call, they said N5 shipped in volume ).
You can cross verify this info from AMD's Earnings call of Q2 as well, and AMD said the first ramp of Console chips happened and increased capacity.
Then after Sep, Huawei chips were stopped and capacity taken over by AMD.

80K wpm is more than half of N7 capacity which is unlikely considering NV, QCM, MTK etc also using the same node not to mention Apple still manufacturing some chips with N7 as of Q4 20.

I own quite a bunch of AMD shares so I dont usually miss TSMC and AMD earnings call.

Right now I think Samsung 5LPE can really help the PC ecosystem, because with Apple using TSMC leading nodes, PC ecosystem is left with scraps on the leading edge nodes, with Intel already out of the race.

Regarding the Fab in US, I was expecting Samsung also to expand. They already have the Austin Fab. Not sure if that changes in the near future.
 
Last edited by a moderator:
tl;dr: NCOV, Intel dropping the ball because Intel and console kiddies fucked the PC market. More news at 11.

Hey, at least I have a reason to wait for a 5950XT on 6nm now!
 
IIRC TSMC is also building a new 5nm fab in the US
This is so no sense. I'm not going to pay 100 bucks over MSRP for a custom card. If AMD want market share this is not the way.

Sigh.
AIB cards traditionally are divided in
- cards at stock frequencies, at MRSP, with reference or almost reference cooling
- mildly OC custom cards, relatively small premium over reference MRSP
- Highly overclocked custom cards with special cooling solutions, with high premium ($100 and more) over MRSP.

This is valid for both AMD and Nvidia cards.

The cards we saw in the tests are ALL of the third category. But, on AIB sites, you can find also the reference models

https://www.sapphiretech.com/en/consumer/21304-01-20g-radeon-rx-6800-xt-16g-gddr6

https://www.powercolor.com/product?id=1604398054

So if the highly overclocked parts' listed price is at 100-150$ over MRSP is perfectly normal, that also means that AIB parts with standard cooling will have a price at or around MSRP.
The problem here is that there is extremely limited supply, so:
- it is difficult to find anything at MSRP
- if AIB receive a GPU, it's very likey they will try to sell it as top-end OC model, because in this moment they will sell everything they maufacture anyway, and top end model means more margins for the AIB

AMD is restocking every now and then reference cards on their site, but supply is not great and will not last forever either.
Nvidia is not in a much better solution, considering they had two months more.

Only solution for all this is availability (not likely to improve before end of December/next year), and not buying at inflated prices. Also, no market share gains if there are no cards.
 
You mean the reference model that the Twitt said won't be restook and that I was talking about? I see.

The AMD branded model will not be restocked after January. There is no statement at all about AIB branded reference models - and you can see those on every AIB site but not listed yet.
 

Oh man. It would explain why AMD appeared relatively co
I know there is not a lot of other solutions, but the ultra dependance on TSMC is kind of worrisome. They can't produce for everyone... Will it be even worse with 5nm ?

Another reason I can see to Samsung, at least partially. AMD,
Level of Detail for BVH has been demonstrated on NV hardware - so it is not a limitation. What is happening in that miles morales screen shot is not level of detail causing t-posing Rather the game code for animation and characters not caring about how they look when the camera is not looking at them. A different "issue" altogether.

Basically, your gameplay systems are more lenient in fidelity and simulation but ray tracing now exposes the cheating.


If you read the modern warfare 2019 ray tracing presentation - they mention how vertex animation grass and foliage is a limitation that they will go around in future products by actually using real mesh representation for grass and leaves and animating them that way.

https://www.activision.com/cdn/research/Raytraced_Shadows_in_Call_of_Duty_Modern_Warfare.pdf

Slide 61: "Alpha testing will most likely have to be replaced with high poly opaque BLASes, but actual performance/mesh complexity balance is yet to be measured."

That would be a positive thing for ray tracing against them as well as the fidelity of such assets. It will be nice to see how these games look once the last gen is dropped and the asset production can catch up to the new techniques.

There definitely are ideas for getting over some limitations. Things can improve, but needing to get all the way back to raster performance on these things for the sake of somewhat better shadows, at a performance gap that will probably never be closed, is why I question the usefulness of all this to begin with.

Virtual shadowmaps, like UE5 uses, can get near the same quality for less coding time and complexity and faster rendering times. To me that's what realtime rendering is about; not "doing what the Hollywood guys do" like raytracing hardware tries. But "replicating the end results" they get a thousand times faster.
 
Back
Top