Discussion RDNA 5 / UDNA (CDNA Next) speculation

[adroc_thurston]

Wasnt there something about RDNA4 not fitted to be used as iGpu

No.
RDNA4.5 iGPs just got cancelled in early 2023.

 

[Tachyonism]

No.
RDNA4.5 iGPs just got cancelled in early 2023.

I guessed they just decided that it wasnt worth the hassle implementing it on client while Intel was so far behind lol. Better wait for RDNA5 next year.

 

[adroc_thurston]

I guessed they just decided that it wasnt worth the hassle implementing it on client while Intel was so far behind lol. Better wait for RDNA5 next year.

This has nothing to do with Intel and everything to do with employee workload and morale.

The new roadmap is built around intercepting major IP/node updates every 8Q or so.
MDS has new CPU, new AIE and new fabric (probably).

 

[Joe NYC]

Ah, you're buttmad AMD is not giving you treats.
Sorry pal, it's a server kingdom.
They don't care about you. They will never care about you.
You will get server scraps and you WILL enjoy them.

AI bubble forever? Because without AI bubble, the server market just goes to lowest cost generic CPU, hyperscalers house brands, freeloading on Arm IP.

Just in case AI is extended to client, AMD is positioning itself perfectly to miss out on it.

Not just miss out. Pissing away a rare leadership position AMD stumbled into - Strix Halo - and go from Hero to Zero.

When product cycle matters most, in charting new territory, AMD takes a 3 year hiatus and gives us "Gorgon Halo".

?
We're in Year 8 of QCOM fumbling the bag.

That's exactly where the problem is. When you have to start comparing AMD with Qualcomm in PC and iGPU space.

AMD set out on this course back when Hector Ruiz was in charge. AMD had decades, and now we are down to comparing AMD with Qualcomm side quest into PC.

 

[adroc_thurston]

Because without AI bubble, the server market just goes to lowest cost generic CPU, hyperscalers house brands, freeloading on Arm IP.

Oh no lmao that's not how it works.
Winning against AMD on TCO is nightmarishly hard.
Venice will show you just HOW hard can it be, really.

Not just miss out. Pissing away a rare leadership position AMD stumbled into - Strix Halo - and go from Hero to Zero.

When product cycle matters most, in charting new territory, AMD takes a 3 year hiatus and gives us "Gorgon Halo".

You being mad over not getting treats does not change the market equilibrium.

That's exactly where the problem is. When you have to start comparing AMD with Qualcomm in PC and iGPU space.

AMD set out on this course back when Hector Ruiz was in charge. AMD had decades, and now we are down to comparing AMD with Qualcomm side quest into PC.

This is very insane to say the same Q AMD gained yet more client rev/unit share.
You really should stop. Intel was selling LNL at cost to stop those share gains and it did *squat*.

 

[Joe NYC]

Oh no lmao that's not how it works.

The "insatiable hunger for compute" at the time of Covid quickly got "satiated" and turned to a hangover. Only to be rescued by the AI bubble last year.

AMD gaining in server was 80% due to Intel just absolutely screwing up. Without Intel screwing up, in a stagnant market, AMD success would have been much less.

Without AI bubble, supply of compute outpaced the demand for traditional cloud services (TAM growth was dead), and lead to the Gravitons etc... Generic, commodity CPU cores.

Now with the server TAM suddenly growing, AMD was (rarely) prepared, and now the benefits are falling in AMD's lap.

Which is why I say AMD needs to be similarly prepared in other markets.

Your argument is that AMD does not need to be prepared elsewhere because AMD got lucky in server CPU, after AI bubble rescued the server CPU market.

You being mad over not getting treats does not change the market equilibrium.

It's not a "treat" to expect AMD to close the big gap behind Apple Mac SoCs. It is a dereliction to not close the gap. It is also dereliction to allow others to catch up and overtake AMD in (i)GPU

This is very insane to say the same Q AMD gained yet more client rev/unit share.
You really should stop. Intel was selling LNL at cost to stop those share gains and it did *squat*.

AMD gaining share is the results of numerous missteps by Intel. The gains can only continue if AMD moves from barely adequate to excellent product offering (since Intel's missteps are coming to an end).

So, you are looking backwards, and I am looking forward.

 

[Bryo4321]

I get it, you're buttmad about client iGPUs.
It's okay pal, you're free to buy Intel.
Oh and it's 3.5 until 2031.

Even for Zen 6 Halo apus?

 

[adroc_thurston]

The "insatiable hunger for compute" at the time of Covid quickly got "satiated" and turned to a hangover

No, DC capex remained fairly flat post 2021.

AMD gaining in server was 80% due to Intel just absolutely screwing up. Without Intel screwing up, in a stagnant market, AMD success would have been much less.

Can you just stop vomiting cope and admit that AMD wins in server? It's not that hard.

Without AI bubble, supply of compute outpaced the demand for traditional cloud services (TAM growth was dead), and lead to the Gravitons etc... Generic, commodity CPU cores.

Graviton is anything but generic, with GV5 being a bespoke 3D stacked part.
Kid, you're wayyyyy outta your depth and should stop.

Your argument is that AMD does not need to be prepared elsewhere because AMD got lucky in server CPU, after AI bubble rescued the server CPU market.

This has nothing to do with luck and everything to do with execution.

It's not a "treat" to expect AMD to close the big gap behind Apple Mac SoCs. It is a dereliction to not close the gap. It is also dereliction to allow others to catch up and overtake AMD in (i)GPU

You're expecting a server company to spend gigaeffort on client parts.
You want treats and AMD gives you none.
You can stay mad or accept the reality.

AMD gaining share is the results of numerous missteps by Intel. The gains can only continue if AMD moves from barely adequate to excellent product offering (since Intel's missteps are coming to an end).

Can you read? Reading letters shouldn't be here.
Ultra 5 LNL at cost mogs every single AMD laptop offering and it didn't stop AMD from gaining share like at all.

So, you are looking backwards, and I am looking forward.

You are throwing a tantrum over AMD not giving you treats.
It's cute but also very pointless.

Even for Zen 6 Halo apus?

They're not Zen6 Halo APUs, but graphics parts from the graphics roadmap. That are also APUs.

 

[Bryo4321]

They're not Zen6 Halo APUs, but graphics parts from the graphics roadmap. That are also APUs.

Gotcha, that makes sense. Thanks for clarification.

 

[Joe NYC]

No, DC capex remained fairly flat post 2021.

Flat 2021-2024, with a period of "indigestion" after the insatiable demand was satiated. It was not going anywhere (and AMD would not have gone anywhere further beyond market share gain without the AI bubble).

Can you just stop vomiting cope and admit that AMD wins in server? It's not that hard.

The winning looks a lot better, only after AI bubble awakened it from a zero sum game.

Graviton is anything but generic, with GV5 being a bespoke 3D stacked part.

Merchant CPU vendors need to continuously provide superior product and value, otherwise, the merchant CPU vendors will be phased out by vertical integration of their most important customers.

Apple, Amazon, Google, Microsoft.

Taking a nap, as Intel did, is what lead to all these in-house CPU / SoC projects.

Which is, why it is so bad for AMD to be napping in SoC / APU, iGPU, Halo.

Halo has an obvious flaw, for it to be used in Local AI, which is poor networking / IO for the task. Correcting that, with some other improvements, such as perhaps a shrink would have been a decent refresh for emerging market (premium / gaming / AI) for mid 2026 before the next Halo arrives 1+ year later.

By taking a 3 year nap, AMD is handing this market to Apple and NVidia. Maybe even Qualcomm and Intel.

This has nothing to do with luck and everything to do with execution.

Execution is preparation. When the market unexpectedly takes off, it can be called foresight or luck.

Since AMD did not see the AI emerging, it was more luck than foresight.

You're expecting a server company to spend gigaeffort on client parts.

That's wrong, if same / similar IP applies to both. It's good ROI if the same IP brings profit from both markets.

You want treats and AMD gives you none.
You can stay mad or accept the reality.

Not treats. Just lessons learned.

What lesson is there to learn from missing out on nearly $700 billion market?
The lesson is not to keep missing out on opportunities which are withing grasp.

https://twitter.com/x/status/2019340117874192469

 

[adroc_thurston]

Flat 2021-2024, with a period of "indigestion" after the insatiable demand was satiated

There was no indigestion for AMD.
If anything ML bubble hurt CPU capex for 2 years.

The winning looks a lot better, only after AI bubble awakened it from a zero sum game.

It put CPU capex into the grave until late last year.

Merchant CPU vendors need to continuously provide superior product and value, otherwise, the merchant CPU vendors will be phased out by vertical integration of their most important customers.

Try outTCOing Venice.

Which is, why it is so bad for AMD to be napping in SoC / APU, iGPU, Halo.

You're throwing a tantrum about not getting treats again.
You're not getting them.

Execution is preparation.

No, execution is execution.

That's wrong, if same / similar IP applies to both. It's good ROI if the same IP brings profit from both markets.

You do understand that IP alone does not build a product.

Not treats. Just lessons learned.

Client slop is treats.
AMD is not giving you client slop.
Server only now.

What lesson is there to learn from missing out on nearly $700 billion market?

That's all server capex.
Has nothing to do with client slop.

 

[ElectricSand]

Client slop is treats.
AMD is not giving you client slop.
Server only now.

Giving up in a market responsible for like 30-40% of their revenue is some loser-ass activity. AMD's not a lil babby company any more. They make a lot of money. Hire more people if their current staff counts can't support a more elaborate product roadmap. Poach people if you must.
Allowing yourself to get beat by Intel in iGP is just embarassing tbh.

 

[adroc_thurston]

Giving up in a market responsible for like 30-40% of their revenue is some loser-ass activity

They're not giving up. Just not doing much there either.
Client rev share will decline into irrelevancy come H1'27 anyway. whatever man.

Hire more people if their current staff counts can't support a more elaborate product roadmap. Poach people if you must.

They did. For data center.
(client graphics is also pretty hiring-rich).

Allowing yourself to get beat by Intel in iGP is just embarassing tbh.

They don't care.
iGP is not a real metric that ever mattered in mobile.
It's CPU perf >= BL >>>>> anything else.

 

[branch_suggestion]

AMD's new mobile roadmap is pretty simple:
MDS1/2/3 - R7 and below replacement for typical notebooks.
MDS1-hi - R9 and partial Fire Range replacement for high CPU perf applications, made to pair with green sticker machines in high ASP markets.
Gator Range - Desktop replacement for extreme gaming laptops.
MDSP/H - GPU driven APUs designed to compete with green sticker machines, should come out roughly 1 year offset from the major CPU update.

Good news is NV probably won't bother releasing GR20x until late 2027/2028 anyway, N2X timing might be important but idk, they've got a lot of work to do.