Discussion RDNA 5 / UDNA (CDNA Next) speculation

[T2098]

Additional revenue implies there's a market for $2k gaming Radeons which is just not true.

It's not the memory prices, it's the PPA that kills it.

There's also a whole bunch of costs that probably don't make sense if you think you're going to sell a very small amount of cards in total; you have to develop, test, and qualify the PCB for the gaming specific SKU, you've got marketing to deal with, and then presumably the cloud version and possibly even the provis version have their own driver stack, so you now have your software engineering costs to bring it into the gaming driver stack, as well as ongoing support and optimization for the same.

AIBs also might not be keen in paying all their own fixed costs for their own PCB development, marketing, support, etc if each AIB might only sell a few hundred.

 

[basix]

No, it's not a price issue at all. AMD loves high ASPs.
Lisa hates mysterymeat programs with ???? results.

Why should the outcome be unknown? AMD can release AT0 after whatever Nvidia releases.
The chip is anyways there and ready to use. They will even build desktop form factor PCBAs and coolers (maybe something beefier than DHE like 5090 / 6000 Pro cooler design).
If AMD sees an opportunity after Nvidias market entry, they can decide relatively late in the process if gaming AT0 will be a thing or not. If there is no opportunity, then no gaming card.

 

[adroc_thurston]

There's also a whole bunch of costs that probably don't make sense if you think you're going to sell a very small amount of cards in total; you have to develop, test, and qualify the PCB for the gaming specific SKU, you've got marketing to deal with, and then presumably the cloud version and possibly even the provis version have their own driver stack, so you now have your software engineering costs to bring it into the gaming driver stack, as well as ongoing support and optimization for the same.

People here really don't understand g2m costs for any product and it's silly.

Why should the outcome be unknown? AMD can release AT0 after whatever Nvidia releases.
The chip is anyways there and ready to use. They will even build desktop form factor PCBAs and coolers (maybe something beefier than DHE like 5090 / 6000 Pro cooler design).

g2m isn't about just making PCBs and stuff.

If AMD sees an opportunity after Nvidias market entry, they can decide relatively late in the process if gaming AT0 will be a thing or not. If there is no opportunity, then no gaming card.

There is no opportunity; never was; never will be. That market is forever closed to AMD.

 

[basix]

g2m isn't about just making PCBs and stuff.

Then please enligthen us. What are we missing?

RTX 6000 Pro is also manufactured by board partners like PNY. So if there is a desktop grade card already available and being produced, the additional work and cost is very, very miniscule compared to everything done up to that point (chip design, PCB design and qualification, cooler design, most driver stuff, manufacturing and logistic channels, ...).

 

[adroc_thurston]

Then please enligthen us. What are we missing?

RTX 6000 Pro is also manufactured by board partners like PNY. So if there is a desktop grade card already available and being produced, the additional work and cost is very, very miniscule compared to everything done up to that point (chip design, PCB design and qualification, cooler design, most driver stuff, manufacturing and logistic channels, ...).

Tons of marketing and channel efforts, a ton of devrel, volume commitments between a bunch of AIBs and a whole other pit of issues caused by trying to push a $2k Radeon into the market.

 

[basix]

You do not do most of that already for all other RDNA5 GPUs? Devrel, marketing, channel and AIB relations?

The only part I see there are volume commitments from AIBs. But if you do not want to handle that: "Manufactured by AMD" is the only option (with a single manufacturer buildig the cards) like AMD will probably already do it for the professional / workstation line of the cards.

I don't know, you somehow stringently seek for a reason that such a card is "unbuildable". But in my opinion 90% if not more of the effort surrounding that chip and card is already done.

 

[Kepler_L2]

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.

 

[jpiniero]

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.

Guaranteed there will be people complaining it should be $599. Not worth the effort.

 

[del42sa]

MLID said AT0 is coming to desktop dGPUs.

must be true then

 

[basix]

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.

That is what I'am talking about. That would be the minimum.

If there is no demand, only little volume for consumers with a single "AMD design". If there is much demand, AMD can increase the volume. And as soon as the volume is there, there is demand for AIB custom designs.

 

[1250]

Sorry to break the mood, but let me talk about the MI455x.
Some memory guys are against custom HBM due to cost.
consulting person(Maybe? for Google Broadcom AMD etc. He knows the capacity of LPDDR.) is from the custom HBM side.
They said they went back standard hbm because of cost issues
Is there a way to break this?
P.s It seems like one stubborn person is in the planning department.

 

[basix]

It depends what "cost issue" exactly means and who you are talking to:

• HBM guys want to sell more HBM and not LPDDR (higher margins)
  • They will happily tell you to take "cheap" standard HBM
• HBM with custom base die will be more expensive than standard HBM, that is a no-brainer (more expensive base die process)
  • The additional cost ("cost issue") must have some other benefit to offset the additonal $$$
• AMD wants to build an overall cheaper and/or more performant system (LPDDR is much cheaper than HBM)
  • Increased cost of HBM with custom base die gets offset by cheaper LPDDR memory and/or overall higher memory capacity (running LLM with bigger batch sizes drives down cost)

 

[1250]

HBM with custom base die will be more expensive than standard HBM, that is a no-brainer (more expensive base die process)

• The additional cost ("cost issue") must have some other benefit to offset the additonal $$$
• AMD wants to build an overall cheaper and/or more performant system (LPDDR is much cheaper than HBM)
  • Increased cost of HBM with custom base die gets offset by cheaper LPDDR memory and/or overall higher memory capacity (running LLM with bigger batch sizes drives down cost)

To be precise, the claim is that the lpddr controller is inside the GPU. LPDDR is connected to the IOD or MID, just like a graphics card.

 

[1250]

It depends what "cost issue" exactly means and who you are talking to:

• HBM guys want to sell more HBM and not LPDDR (higher margins)
  • They will happily tell you to take "cheap" standard HBM
• HBM with custom base die will be more expensive than standard HBM, that is a no-brainer (more expensive base die process)
  • The additional cost ("cost issue") must have some other benefit to offset the additonal $$$
• AMD wants to build an overall cheaper and/or more performant system (LPDDR is much cheaper than HBM)
  • Increased cost of HBM with custom base die gets offset by cheaper LPDDR memory and/or overall higher memory capacity (running LLM with bigger batch sizes drives down cost)

The claimed development cost is in the hundreds of millions of dollars. This includes the cost of designing and testing the three base die and core die, etc., and the mask is about 50 million.

 

[Momoka_]

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.

So RDNA5 is still trailing the RTX 60-series in terms of release?

 

[1250]

The compromise is that the Samsung combination and the TSMC combination competed and the Samsung combination (custom B-die) was defeated. What do you think? MI455X
tsmc MID+bigger I/O or Samsung MID + I/O + C-HBM

 

[marees]

So RDNA5 is still trailing the RTX 60-series in terms of release?

unless they have a flagship product they will have to trail

the strong do what they can
the weak suffer what they must

 

[1250]

Samsung combination (custom B-die) was defeated. What do you think? MI455X

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?

 

[ETI4711]

In November 2024 Marvel announced to do custom do custom HBM and presented that:

https://twitter.com/x/status/1866546802289414631

The advantages are obvious. Please see the discussion in the tweet.

Nvidia is producing millions of data center GPUs per year. AMD talks about tenth of billions revenue in 2027.
Can they afford to ignore such an advantage?

They say to wait for Nvidia's c-hbm4e, which won't be available until 2028.

There is other news available. But you have to thrust the source. As always.

NVIDIA and AMD Customize HBM4 Memory for Next-Generation AI Accelerators

NVIDIA's and AMD's next-generation accelerators will definitely utilize a next-generation HBM4 memory, too. However, interesting news, confirmed by SemiAnalysis, is that these accelerators will utilize a customized HBM memory for the first time. In an HBM stack, consisting of stacked DRAM dies...

www.techpowerup.com

There's nothing I can say to convince him. I don't know much about this community. Top secret?

There is nothing secret about custom HBM Samsung , SK Hynix and others talk about it.

SK Hynix reportedly clinches custom AI memory orders from Nvidia, Microsoft, Broadcom

SK Hynix has reportedly secured customized high-bandwidth memory (HBM) orders from Nvidia, Microsoft, and Broadcom, pulling ahead of domestic rival Samsung Electronics in the race to supply AI-optimized memory solutions, according to reports by and .

www.digitimes.com

https://www.edn.com/collaboration-delivers-custom-hbm-solutions/

How Custom HBM is Shaping AI Chip Technology

Expert insights and analysis from industry insiders, with both organic, personality-driven content and strategic client partnerships, we’re leading the conversations on enterprise technology’s evolution.

www.sixfivemedia.com

High Bandwidth Memory (HBM): Customization vs. Standardization | Synopsys

Discover how High Bandwidth Memory (HBM) is shaping AI's future. Explore the debate on customization vs. standardization and the impact on innovation.

www.synopsys.com

SK Hynix plans 30x improvement on HBM with custom variants

Ryu Seong-su, vice president of SK Hynix, has announced plans for his company to develop products with 30x the performance of HBM DRAM.

www.eenewseurope.com

https://www.eetimes.com/hbm-innovation-outpaces-standards-development/

But Nvidia and AMD don't talk much about unreleased products. And even after release they do not talk to much, ...

But if you are lucky Nvidia is talking about custom HBM at the GDC 2026. Or AMD is showing how the LPDDR is connected to the MI450.