Discussion RDNA 5 / UDNA (CDNA Next) speculation

[adroc_thurston]

Meta is not getting the fully enabled chip?

They are. Different flexIO splits, though.

 

[1250]

They are. Different flexIO splits, though.

What is that kid talking about when he was the first to spread the news about LPDDR?
custom HBM semianalysis. 819gb/s

 

[adroc_thurston]

What is that kid talking about when he was the first to spread the news about LPDDR?

who

 

[1250]

who

Semianalysis
If it weren't for him, I wouldn't have come this far
BTW, will the memory-bad guys notice this year?

 

[CINEBENCH_ACCELERATOR]

There are reports of Sony contemplating delaying PS6 to 2028 or even 2029.

2028 seems like a no brainer given memory prices, and in any case they will benefit from having as large PS5 installment base as possible.

Now N3 was released in 2023, so how silly it would be to use N3 design in 2029 for a console meant to last like 7+ years, that's like Switch 1 territory rather than next gen.

Would be really pathetic if they don't bite the bullet and make new design for N2.

Very different things.

N3 will remain strong in 2030s. Switch 1/2 use Ancient Litho & Bad Chip. Terrible combo!

But RDNA5 will be N2 right?

 

[adroc_thurston]

But RDNA5 will be N2 right

They're all N3p.

 

[gdansk]

They're all N3p.

only RDNA3¾ on n3c?

 

[adroc_thurston]

only RDNA3¾ on n3c?

Nothing is on N3c.

 

[gdansk]

Nothing is on N3c.

Oh good

 

[Claudiovict]

They're all N3p.

N3X wouldnt be better for frequency? Given AMD didnt reach its targets in the last generations, going with N3X wouldnt be better?

 

[adroc_thurston]

N3X wouldnt be better for frequency? Given AMD didnt reach its targets in the last generations, going with N3X wouldnt be better?

Not applicable to GPUs.

 

[dangerman1337]

Not applicable to GPUs.

N3X can't be used for GPUs? Wonder why?

 

[adroc_thurston]

N3X can't be used for GPUs? Wonder why?

As in 2% speed at more money is hardly relevant for GPUs.

 

[1250]

I don't know if I should be saying this here, but I'm curious about what AMD did with the LPDDR5x price (they originally wanted to raise it by 60% for Apple). A little LPDDR info: LPDDR6 D1B, D1C starting at 24GB

 

[1250]

Meanwhile, there is talk that NVIDIA's c-hbm4e is n3p
BTW, helios rack same(kyber, It's wide though)

 

[Joe NYC]

Unless AT0 is cancelled entirely it will 100% release to consumers, even if it's a token release with no real volume like the Radeon VII.

Makes a lot of sense. You never know what might come along to boost the demand significantly. There was crypto at one point, and then there is AI, which can take a lot of unexpected turns.

Today's NVidia results:
- Gaming + AI: $3.7 billion
- Professional Visualization: $1.3 billion

Total: $5 billion
(which is nearly the same as AMD company-wide revenue, including server, AI GPU, Xilinx FGPA, client PCs).

NVidia is no longer hiding that the "gaming" dGPUs are used for AI etc. Pure gaming, of the $5 billion dGPU sales is probably 1/2 or less.

It would be foolish for AMD to spend all the money to develop product such as AT0 and not offer it as an option to this multi-billion dollar market.

 

[adroc_thurston]

It would be foolish for AMD to spend all the money to develop product such as AT0 and not offer it as an option to this multi-billion dollar market

There is no market for $2k Radeons.
It'll move maybe 5k units *in total*. Probably less.

 

[Joe NYC]

They say to wait for Nvidia's c-hbm4e, which won't be available until 2028. There's nothing I can say to convince him. I don't know much about this community. Top secret?
2028 is too long. How many months do I have to wait to see results?

BTW, at which point to you think they will start making HBM with hybrid bond?

It seems that at some point, the plan was regular HBM4 and then everyone dropped the ball on it. Will it be HBM4e now?

 

[Joe NYC]

There is no market for $2k Radeons.
It'll move maybe 5k units *in total*. Probably less.

But you are ignoring the market for AI Pro Max+ eon

 

[adroc_thurston]

But you are ignoring the market for AI Pro Max+ eon

That's all Nvidia.
There is no market for a $2k Radeon. Never was, never will be.
Which is why any attempt at a $2k Radeon is always killed during a program review.

 

[1250]

BTW, at which point to you think they will start making HBM with hybrid bond?

It seems that at some point, the plan was regular HBM4 and then everyone dropped the ball on it. Will it be HBM4e now?

They say they are currently considering a wafer-to-wafer bonding method (W2W). Micron has a sample of 4x5 bonded 4 wafers and 5 wafers bonded using the TC-NCF method (20Hi).

 

[Joe NYC]

That's all Nvidia.
There is no market for a $2k Radeon. Never was, never will be.
Which is why any attempt at a $2k Radeon is always killed during a program review.

Even Intel is trying to release some sort of high end GPU card pretending to be AI card. It will be using LPDDR5x.

Why not use some versions of AT0 and release it in a way that it could be used for local AI.

It would be foolish for AMD to not have a product offering, or even better, multiple products from Medusa Halo to AT0 for potential use as local AI.

I think programmers might like a dedicated local AI system ...

 

[Claudiovict]

There is no market for $2k Radeons.
It'll move maybe 5k units *in total*. Probably less.

It has to be 2k? 9070 XT is 357mm² with MSRP of $600, if AMD double the chip to 714mm², they couldnt charge $1500? (or 1600, 1650?)

 

[Joe NYC]

They are. Different flexIO splits, though.

Is the flexIO the million PCIe6 lanes that Mi455 will have, to be used defiantly for Meta?

 

[Joe NYC]

I don't know if I should be saying this here, but I'm curious about what AMD did with the LPDDR5x price (they originally wanted to raise it by 60% for Apple). A little LPDDR info: LPDDR6 D1B, D1C starting at 24GB d1d good die success

It seems that the LPDDR5x is optional, so whoever is willing to pay...