AMD: RDNA 3 Speculation, Rumours and Discussion

TBF I think a entry level card that is 8GB with mid range being firmly 12GB in 2023 for RDNA 3/RTX 40 would be fine. I mean if we end up with entry level cards (<200 USD) that are 3070 Ti levels of performance that'd be great for 1080P gaming. I mean 1080p is going to be with us a long time for many lower end setups or those who mainly want high frame rates without breaking the bank. If we have Nvidia say release a AD107/108 that has 3070 Ti performance but has it as 8GB and one played 1080p gaming with "good enough" levels of framerate (120+ in some games, 60+ in heavier next-gen titles) you wouldn't need more than that.

 

Will not exist.

 

I think you are rather too quick to state that. People always overfit current trends. I think you are doing precisely that.

There is currently no low-end because AMD and nV are both consuming all available fab capacity without exhausting the demand for more expensive, higher-margin parts. This situation will not be permanent. Near the beginning of the fab shortage, massive investments in fab capacity were published. In a few short years, we will go from a shortage of fab capacity to a glut of it. In that environment, there absolutely will be low-end parts, not just at $200 but also at $100.

They might not be made on the latest and greatest process, so especially in power efficiency might not be as near the top-end as past generations, but they will exist.

 

Rebrands are all you're getting and that's from NV.
AMD? Too bad!

Oh it is permament for AMD as long as Intel exists.
There's never quite enough stuff when it comes to voring Intel.

 

I'd argue that there is: it's filled with previous gen parts going back as far as to Pascal and Polaris. 3050 is also a thing on Nvidia side.

 

I just need a cheapo hdmi 2.1 card for game streaming to the tv at 120 fps. Niche use case for sure so I’ll be waiting a while.

 

You really haven't been paying attention to the foundry expansions. Past 2024 or so, TSMC will be able to serve all the market currently served by Intel, on top of all their current customers, with healthy growth too.

Then add all the other fabs that are coming online. There will be not-quite-high-end-but-almost fab capacity that will be underutilized. Literally making nothing.

Do not overextrapolate from current trends.

 

Of course I did.
Most of foundry cap ramp is trailing edge lol.

dawg I literally made money by doing just that.

 

Probably but I already have an HTPC. Just need to upgrade the GPU. Paid $150 for the 1650 super in it now.

 

TSMC is curently manufacturing almost nothing for Nvidia, very low volumes for Intel and covering like half the silicon used by AMD. In the near future, Nvidia will be manufacturing both Lovelace and Hopper at TSMC, Intel will manufacture almost all GPUs and many tiles at TSMC and AMD will move almost all the chiplets and lower-mainstream products (currently manufactured by GlobalFoundries) to TSMC. It's really hard to imagine that TSMC will (still) have available capacity, which could be used for low-end products.

 

Current expansion / investment plans for TSMC are mainly for

1. Bringing up volume of N5 up to N7 levels, which according to the roadmap below is till end of 2023. Including American fab
https://fuse.wikichip.org/news/5760/tsmc-2021-foundry-update-foundry-roadmap/
According to this data, N5 is less than half N7 output which is definitely nowhere near enough.
Already booked all the way to end of 2022 according to earnings report.
Currently Fab18 is doing N5, It is like Phase 3 of expansion going on I believe.

2. Ramping up N3
Don't know where this is, probably Fab18 due to more recent EUV equipment installation there.

3. Building new fabs for N2 and beyond
Fab 20

I guess N7/N6 will take over the role of 16FF, once it becomes cheap enough.
Fab 14 and 15 are doing some form of N7/N6 (excluding Fab12A,12B which were doing some limited production and research)
N7 family capacity would no longer be expanded after end of 2021 as per TSMC roadmap

The leading edge nodes would continue to remain dear for the foreseeable future.
Post 2024 It is GAAFET/MBCFET era. Don't know what it will bring. If it is as good as touted by everyone, everybody will be jumping on it.

 

There we have it! This basically proofs you must be wrong

 

TSMC 5nm allocation on 2021 (data by Counterpoint):

Edit: For reference, 5nm production will account for 5% of 12-inch wafers this year at TSMC

Edit2: And below, 2021 TSMC 7nm production allocation:

 

Spot the bs.

 

Looks like the boxed era comes to an end sooner than expected. Obviously streaming is the best solution for chip shortage?
I'll need some time to swallow this down. But i see some upsides. Like >1TB games, and massive multiplayer becoming much easier. :/

 

I'm not sure if putting more silicon per user-hour played is a good idea for that.
Speedy NICs ain't free and neither are the ToR switches!

 

I have a solution to reduce user hours too: 'Insert coin.'
We are all fucked.

 

Thank you Spencer-san, I will buy your fucking sub sooner or later.

See?
Now that's the appropriate attitude.

Jokes aside, xtor cost scaling biting the dust also kills all and every GPU value prop since more GPU goodness == more xtors.

 

Those are averages for 2021, which is in its second half. The charts tell us more about past than about future.