Discussion RDNA 5 / UDNA (CDNA Next) speculation

[SolidQ]

2026 is too soon for any kind of RDNA 5 hardware right ????

Technically RDNA4/RTX 5xxx is 2024 hardware, they just delayed. So standard 2 year cycle is should be in 2026

 

[marees]

Technically RDNA4/RTX 5xxx is 2024 hardware, they just delayed. So standard 2 year cycle is should be in 2026

maybe microsoft can take the medusa premium & release it (via the board partners) as the series s2 ???

 

[adroc_thurston]

maybe microsoft can take the medusa premium & release it (via the board partners) as the series s2 ???

you really should stop.

 

[Mopetar]

They can, in theory, roll out multiple official Xbox devices made by various partners using whatever hardware and do this on an annual basis.

So it's a PC?

 

[gdansk]

So it's a PC?

No, PCs actually sell. It would be an Xbox. It wouldn't sell.

 

[Win2012R2]

nope , if Zen was a consumer focused core you would likely see a very different cache setup and in a given xtop/mm^2 budget less Load/store/TLB and more execution resources.

And who'll pay for this extra R&D, masks, diff inventory in stock - you? It's a clever solution by AMD that is obviously tilted a bit towards server market because that's the main earner.

9800X3D had custom consumer level IOD, so that's bona fide consumer part.

But the most important obvious thing is this - consumer is whatever AMD says by virtue of creating SKU targeted as consumer market, same as AD102 is reused for enterprise and consumers (with money)

 

[Vikv1918]

"Full bore Windows" meaning Steam as well?

 

[Claudiovict]

If this confirms to be true, and RDNA 5 turns out as a good product, AMD can get some goods if they launch it this year (H2 according to Kepler and Adroc)

 

[adroc_thurston]

AMD can get some goods if they launch it this year

it's 2027

 

[Kepler_L2]

If this confirms to be true, and RDNA 5 turns out as a good product, AMD can get some goods if they launch it this year (H2 according to Kepler and Adroc)

View attachment 136429

RDNA5 has been 2027 for a while, and they can't launch before NVIDIA anyway

 

[Saylick]

RDNA5 has been 2027 for a while, and they can't launch before NVIDIA anyway

Why's that? Is it the "Nvidia - $50" strategy?

 

[adroc_thurston]

Is it the "Nvidia - $50" strategy?

It's not a strategy, they don't have a halo part to build the stack around.

 

[Claudiovict]

RDNA5 has been 2027 for a while, and they can't launch before NVIDIA anyway

oh hell, thats sad

 

[inquiss]

oh hell, thats sad

And yes it's an obvious consequence of the market preference

 

[branch_suggestion]

Both vendors are also waiting for the second gen GDDR7 ramp, which is a very necessary membw bump.

 

[adroc_thurston]

Both vendors are also waiting for the second gen GDDR7 ramp, which is a very necessary membw bump.

super not necessary given cost per xtor stalled shader growth into nothingness

 

[branch_suggestion]

super not necessary given cost per xtor stalled shader growth into nothingness

Faster memory does save a bit of shoreline at the cost of the memory being more expensive per GB, also slop needs the bandwidth.

 

[adroc_thurston]

Faster memory does save a bit of shoreline at the cost of the memory being more expensive per GB

Capacity would still define the shoreline tho.

also slop needs the bandwidth.

meh slopmachines are their own parts.
Client is stuck with little to no shader growth due to cost per xtor being flat or up on the bleeding edge.

 

[reaperrr3]

RDNA5 has been 2027 for a while, and they can't launch before NVIDIA anyway

It's not a strategy, they don't have a halo part to build the stack around.

I have no doubt that you're both (unfortunately) right on this, but I honestly find AMD's way of thinking extremely frustrating.

Why not aim for getting to the market at least half a year before Nvidia, enjoy the "new and shiny" early adopter rush, and then adjust prices when NV launches their stuff?

I mean, even if IPC improvements and PT improvements of RDNA5 fell somewhat short of expectations, I'd still assume AT2 will beat the 5080 in most games, which costs north of $1K, so NV would not have any immediate counter that wouldn't kill their own juicy margins, so I doubt NV would even bother responding in a meaningful way until Rubin launches.

Anyway, the whole situation makes you wonder whether AMD wouldn't have been better off investing some resources in an "N48+50%" N49 afterall, even if that had launched a year later (aka around now).
I mean, that'd probably have fit into 500mm2, possibly been pin-compatible with N31 boards, demolished the 5080, and ~tied the 4090, so they could've sold a 500mm2 N4 chip for >=2x the N48 price, and probably even at sizeable volume.
Given how big GB102 is and how little NV cares about the gaming market now, I'm not convinced NV would've even bothered with a 5080Ti, at most they might've resumed 4090 production at limited volume or pulled in the 5080 Super, and shrugged off any marketshare wins AMD might've accomplished.

 

[Kepler_L2]

I have no doubt that you're both (unfortunately) right on this, but I honestly find AMD's way of thinking extremely frustrating.

Why not aim for getting to the market at least half a year before Nvidia, enjoy the "new and shiny" early adopter rush, and then adjust prices when NV launches their stuff?

I mean, even if IPC improvements and PT improvements of RDNA5 fell somewhat short of expectations, I'd still assume AT2 will beat the 5080 in most games, which costs north of $1K, so NV would not have any immediate counter that wouldn't kill their own juicy margins, so I doubt NV would even bother responding in a meaningful way until Rubin launches.

Anyway, the whole situation makes you wonder whether AMD wouldn't have been better off investing some resources in an "N48+50%" N49 afterall, even if that had launched a year later (aka around now).
I mean, that'd probably have fit into 500mm2, possibly been pin-compatible with N31 boards, demolished the 5080, and ~tied the 4090, so they could've sold a 500mm2 N4 chip for >=2x the N48 price, and probably even at sizeable volume.
Given how big GB102 is and how little NV cares about the gaming market now, I'm not convinced NV would've even bothered with a 5080Ti, at most they might've resumed 4090 production at limited volume or pulled in the 5080 Super, and shrugged off any marketshare wins AMD might've accomplished.

NVIDIA margins are massive and they can counter almost anything with a price drop. What AMD needs is something that is in a difference performance bracket altogether, like the 9800X3D is.

 

[del42sa]

I have no doubt that you're both (unfortunately) right on this, but I honestly find AMD's way of thinking extremely frustrating.

"AMD never miss an opportunity to miss an opportunity"

 

[Joe NYC]

NVIDIA margins are massive and they can counter almost anything with a price drop. What AMD needs is something that is in a difference performance bracket altogether, like the 9800X3D is.

If @adroc_thurston is correct about Mi400 using base die for L2, AMD may already have something.

But even if it is a promising prospect, it will miss the boat for RDNA5.

 

[Joe NYC]

I have no doubt that you're both (unfortunately) right on this, but I honestly find AMD's way of thinking extremely frustrating.

Why not aim for getting to the market at least half a year before Nvidia, enjoy the "new and shiny" early adopter rush, and then adjust prices when NV launches their stuff?

The stars were just not aligned for 2026 launch, and they are all aligned for 2027 launch.

RDNA5 is probably the most ambitious effort in AMD history, to have chiplet based GPU lego pieces that will span all the markets from console, handheld, APU, chiplet, iGPU, dGPU. I don't think it is realistic to expect this effort to take less time than, say RDNA4 which is very limited compered to RDNA5.

 

[marees]

"AMD never miss an opportunity to miss an opportunity"

i asked grok if AMD/Xbox/Microsoft can release the medusa premium this year.

it said no. FP10 socket will not be ready

 

[Joe NYC]

i asked grok if AMD/Xbox/Microsoft can release the medusa premium this year.

it said no. FP10 socket will not be ready

IIRC, the mobile sockets are:
FP10 - universal AMD socket, 2 x 64 bit memory channels, DDR5 / LPDDR5
FP11 - Strix Halo - 4 x 64 bit memory channels, LPDDR5
FP12 - Medusa Halo - 4 x 96 bit = 384 bit LPDDR6

So, it may be socket FP12 that will not be released yet. But there is extremely little info on this out there for LLM to "train on" to get the right answer.