• We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.

Discussion RDNA 5 / UDNA (CDNA Next) speculation

Page 115 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
The Koreans are memory-focused and will worship the company which has all the highest margin memory chips reserved.
Oh well. 5% market share in that market isn't a dead end.
 
gdansk said:
The Koreans are memory-focused and will worship the company which has all the highest margin memory chips reserved.
Oh well. 5% market share in that market isn't a dead end.
Click to expand...
They like to call Radeon a shark sucker(remora?) Idk if from another country
 
poke01 said:
I’m excited for the Xbox pc. I like the concept
Click to expand...
It is a nice paradigm yeah.
Fact is APUs are the way forward outta the ATX dGPU hellscape we are currently in.
Just need a socketable APU platform that supports halo tier gfx configs, basically a mini OAM.
Issue is memory flexibility, you need MoP of some sort but for high capacity you also want a second tier of socketable memory.

Like just 512b G7 is probably fine for gaming and general OS usage but you sorely lack capacity for heavier CPU/GPGPU workloads.

Best balance is probably G7 MoP+LPCAMM2, two distinct memory tiers with GMD having GDDR PHY, CPU/SOC having LPDDR PHY.
HBM and DDR6 are both off limits outside of DC.
You can also use good old DDR5 during the transition period, and the APU socket for lower tier products can use LPDDR on package.

Key thing is you do want enough D2D bandwidth so the GPU can harness the second memory tier at full speed, with the CPU getting enough of the first memory tier to saturate it.

The key issue with this idea outside of being a completely new paradigm vs a huge incumbent install base and supply chain is... it ain't cheap.
Like ATX exists till this day because it is cheaper to have a crummy CPU socket and an isolated GPU on a cost optimised board with only PCIe to unify them.
For games, Work Graphs takes enough burden off the CPU that the PCIe bottleneck is a non factor and thus this platform isn't really necessary.

Now you could just straight up copy Apple as mobile big APUs already are, but that does not scale to the performance levels required to compete with xx90 discrete NV stuff.

The fact is though that only NV can justify shipping such parts in the current PC ecosystem, the client dGPU market is only 20ish billion TAM.
AMD would need to take multiple big gambles over the best part of a decade to maybe get to 40% of that market, that is little more than $2b/q when AMD client and DC already make more than that.
Also margins wouldn't even be that good assuming NV fight hard to keep share.

The goal for AMD and well Intel is to move the client graphics TAM more towards the overall client TAM as a whole, which is much larger.
Specifically away from the TAM that NV currently has 95% of, even if that market doesn't shrink, if you can grow the overall market meaningfully that achieves the same goal.

So as it currently stands, the only way for AMD to ship chungus GPUs in client justifiably would be in the form of 600w+ x86 APUs or as limited run prosumer MBA discrete cards shared with another project (Radeon VII), which is probably the best we can hope for with AT0.

Realistically there needs to be a killer app that sucks with split CPU+dGPU but works amazing with an APU to kickstart such a move. Maybe Helix does take some share away from green sticker people?
After all with each RDNA iteration they can release a new, upgraded Nextbox with first class software support from MS.
But will they ever release such a chungus extreme version to take on flagship NV?

In summary though, AMD graphics are going to be fine, mobile strategy should work out fine, consoles and supporting infra are reliable demand with even stronger API framing specifically around Radeon IP as the first class citizen, I look forward to DX13 and GFX13...
Unlucky for some?

/rant
 
marees said:
seems to me this will increase TSMC prices & consequently also GPU prices

https://twitter.com/x/status/2028915408913207591
Click to expand...

Here is what is likely going to happen:
1. Intel has about 75% unit share of server market:
J

News Post in thread 'AMD 4Q25 Earnings'

Mercury numbers show AMD had a great Q4. Even gained share in mobile.

https://twitter.com/x/status/2021695106198163569
  • Like
  • Wow
2. Of the chips Intel sells, ~75% are based on Intel 7 process node, which is already in short supply. In other words, 56% of server chips are made on Intel 7 process node.
3. Either main, or the only Intel fab that makes Intel 7 is based in Israel.
4. The life expectancy of that Israeli fab is ~1 month

So, if you thought chips were tight, get ready for the mother of all server CPU shortages, as 56% of server CPU supply is vaporized.

AMD can just as well redirect as much N4 capacity to Turin as it can, if TSMC can't increase AMD N4 capacity allotment.
 
branch_suggestion said:
Best balance is probably G7 MoP+LPCAMM2, two distinct memory tiers with GMD having GDDR PHY, CPU/SOC having LPDDR PHY.
HBM and DDR6 are both off limits outside of DC.
You can also use good old DDR5 during the transition period, and the APU socket for lower tier products can use LPDDR on package
Click to expand...

That's way too complicated and expensive. Magnus is going to be well north of a grand as it is.
 
jpiniero said:
That's way too complicated and expensive. Magnus is going to be well north of a grand as it is.
Click to expand...
To be fair it is kinda a mixed use HEDT proposal in the $5k+ range.
For client gaming and typical workloads, 64GB of GDDR7 alone is plenty for a halo APU in the $3k range.
adroc_thurston said:
more tiers is more tears.
Click to expand...
Well you could do like, 768b LP6 MoP but you may need a bit of extra L2.
And well hybrid bonding outside of Zen CCDs and Instinct is not realistic.
In GPGPU lands this exact thing is happening for KVCache memes.
 
Tachyonism said:
Does Sony enforce PSSR requirement on their platform?
Click to expand...
We should know more about the exact PS6 SDK changes and what they want from devs sometime next year.
But I would say that PSSR will be highly recommended along with their other forks of AMD's standard DX native software efforts.
Tachyonism said:
And if they do, does this have any effect on FSR adoption for PC games?
Click to expand...
Well the next Xbox is a PC, and will run all of the Playstation titles that have been ported to PC, which is naturally why Sony ain't porting anymore titles to PC ever.
So no, but what is interesting is that anything built for Helix is built for any DX13 compliant GPU, but anything built for Helix will at the very least use FSR5 but ideally the whole FSR Diamond suite as necessary.
This means Helix native games are GFX13 native games, all RDNA5 parts will benefit and this time for real, no more console ports that are GameWorksified to hell.
So NV is gonna have to convince people to build for NV specifically with only the baseline features to support AMD/Xbox.
As for game engines well Sony are big Epic investors so Unreal should support them well, and therefore GFX13 well.

Xbox until now has only faced Playstation, now they face NV accelerated PCs. The PS6 is really all alone, though the handheld does compete with Switch 2 so Nintendo isn't alone either.
At the end of the day this is basically MS hedging to stop NV from completely monopolising a whole chunk of PCs to themselves.
AMD alone cannot bring NV back to earth, but with the help of Satay Nutella, there is hope.

Valve and Intel are in the corner sulking, Qualcomm is is maximum copium mode as they are like the only major semicon tanking in the biggest bubble ever. Khronos remains a hobby collective, maybe Work Graphs will do what Mantle couldn't.

Though I do hope for the sake of consumers that the capabilities of DX13 are enough that 3rd parties can make good enough vendor agnostic ML solutions for graphics to hedge against extreme vendor specific solutions, which naturally all IHVs have moved to.
 
adroc_thurston said:
Yes they can, it's just expensive and basically a casino.
Click to expand...
Of course they can and are most certainly trying elsewhere.
But as you said it, they don't want to try to win in the traditional sense since it is expensive and full of dumb luck for smaller returns than elsewhere.
So you know, partnerships(tm) is the way to go.

More accurately I should say that AMD alone won't bother bringing NV back to earth, but will bother is there are enough interested partners tagging along.
But you simply cannot pull it off without winning at the top, that is the one commitment currently missing from their plans for now.
 
You must log in or register to reply here.

TRENDING THREADS

Back
Top