Discussion RDNA 5 / UDNA (CDNA Next) speculation

[ToTTenTranz]

I think you need to stop being concerned about the semantics of that word when people have explained to you multiples times what is meant by it.

I think you need to stop trying to make up excuses and convoluted explanations for what @adroc_thurston writes.

AMD do not care to specifically compete with Nvidia in add in card sales in any serious way. That's it.

What's "in any serious way"? What does "compete in a non-serious way" even mean? That AMD is launching products on the market, spending millions on marketing, dev relations, driver development, open source tools for games, SKU development, chip development, chip fab, chip packaging, AIB relations, etc. etc. just for lulz because they're fine if they don't even make any money nor gain marketshare from it?

Do you know how ridiculous this sounds?

 

[inquiss]

I think you need to stop trying to make up excuses and convoluted explanations for what @adroc_thurston writes.



What's "in any serious way"? What does "compete in a non-serious way" even mean? That AMD is launching products on the market, spending millions on marketing, dev relations, driver development, open source tools for games, SKU development, chip development, chip fab, chip packaging, AIB relations, etc. etc. just for lulz because they're fine if they don't even make any money nor gain marketshare from it?

Do you know how ridiculous this sounds?

I don't have to make it sound ridiculous, you just need to look at the product cancellations and how they design their SKUs. They need consistently new IP to make their APUs and consoles work. They've stopped competing seriously with Nvidia in add in board sales and have moved those resources elsewhere.

 

[ToTTenTranz]

I don't have to make it sound ridiculous, you just need to look at the product cancellations and how they design their SKUs.

For RDNA5 I'm seeing consumer dGPU SKUs based off at least 4 different ASICs: AT0, AT2, AT3 and AT4. RDNA3 got 3 chips (N31,N32, N33) from which they launched 10 different SKUs for the consumer AIB market alone. With RDNA5 we might be looking at more than this. Even an AT4 dGPU with LPDDR6 might be very successful as a gaming GPU with moderate PS5-like performance that also doubles down as a decent solution for local AI models and agents due to access to enough DRAM.

AT0 is a much bigger chip than AMD ever released in the consumer / prosumer space. Just because some of these GPU chiplets can attach to CPU chiplets to make up laptop/handheld/console friendly solutions simply means AMD is optimizing their hardware stack through modularity which is good for ROI.

They've stopped competing seriously with Nvidia in add in board sales and have moved those resources elsewhere.

The amount of investment needed to put a consumer dGPU out there from conception to the shelves means it's a serious effort from day one. Calling it "non serious competition" is a nonsensical rhetoric.

 

[adroc_thurston]

For RDNA5 I'm seeing consumer dGPU SKUs based off at least 4 different ASICs: AT0, AT2, AT3 and AT4.

Not a single one of those is a consumer dGPU.

RDNA3 got 3 chips (N31,N32, N33)

And RDNA5 has 0 (zero) chips.

AT0 is a much bigger chip than AMD ever released in the consumer / prosumer space

Have you uhhhh seen the die area?

The amount of investment needed to put a consumer dGPU out there from conception to the shelves

None of those are consumer dGPUs.
They're throwaway bins from other markets.
It's why AT1 is dead. It doesn't have another market.

 

[ToTTenTranz]

Not a single one of those is a consumer dGPU.

It'll start with 4 consumer dGPU SKUs out of the top 2 chips alone. Refreshes along 2027-29 are probably going to bring higher performing versions of AT0 into the mix, akin to 3090 -> 3090 Ti.

And RDNA5 has 0 (zero) chips.

RDNA5 is cancelled now?

None of those are consumer dGPUs.

There will be plenty of consumer dGPU SKUs. You need to cope a bit better about this.

 

[adroc_thurston]

It'll start with 4 consumer dGPU SKUs out of the top 2 chips alone

You know well enough that die rejects from other markets do not make a consumer dGPU.

RDNA5 is cancelled now?

Discrete one? Hasn't been a thing for years.

There will be plenty of consumer dGPU SKUs

Can I put a strix1 on a PCIe card and claim it's your plenty of consumer dGPU SKUs?

 

[ToTTenTranz]

You know well enough that die rejects from other markets do not make a consumer dGPU.

We know well enough that a consumer dGPU is defined by the fact that a consumer SKU is being sold to the consumer as a final product.

Discrete one? Hasn't been a thing for years.

AT0/AT2/AT3/AT4 + MID = RDNA5 dGPU.

Has been a thing for years.

Can I put a strix1 on a PCIe card and claim it's your plenty of consumer dGPU SKUs?

If AMD and an AIB OEM made such a SKU, it would obviously be a consumer dGPU SKU.

 

[adroc_thurston]

We know well enough that a consumer dGPU is defined by the fact that a consumer SKU is being sold to the consumer as a final product.

No it's not and you know it.
Is Vega20 in Radeon7 a consumer GPU?

AT0/AT2/AT3/AT4 + MID = RDNA5 dGPU.

None of those are client discrete graphics.

If AMD and an AIB OEM made such a SKU, it would obviously be a consumer dGPU SKU.

Really stretching the definition here.

 

[techjunkie123]

IMO what matters is competitive performance and pricing in the client space. Doesn't matter what the part was originally designed for. The X3D chips were prototyped for server, not for gaming.

Whether that happens remains to be seen.

 

[ToTTenTranz]

No it's not and you know it.

SKU target is defined by who sells to whom, not the designer of each of its components.
Dual-use has been a thing for decades, especially in GPUs.

Is Vega20 in Radeon7 a consumer GPU?

Let's ask AMD:

None of those are client discrete graphics.

If they are sold as client discrete graphics (which they will be), they are client discrete graphics.

Really stretching the definition here.

Definition of SKU never changed and I'm simply applying it appropriately.

 

[adroc_thurston]

SKU target is defined by who sells to whom, not the designer of each of its components.

Configs are defined by the primary market they sit in.

Let's ask AMD

AMD says everything with DPFP at half rate is HPC. Too bad!

If they are sold as client discrete graphics (which they will be), they are client discrete graphics.

Really stretching the definition here.
Is H100 (with its single gfx-enabled TPC) a client GPU?

 

[Bigos]

Can you create "semantics" thread instead?

 

[Vikv1918]

I don't understand the "we're pivoting away from dGPUs" move at precisely the point when AMD has their best dGPU arch relative to nvidia in the past 15 years. Its not like theyre hurting for cash anything. What else are they gonna do with their $ billions of AI profits?

But the way they're treating the software side is what convinces me of this "we don't care about dGPU" theory. I can't think of any reason why they refuse to bring FSR4 to older gens other than "we care so little about this market we're gonna fk over loyal customers lololol"

I don't think there is any innocent explanation for the retreat from dGPU. My headcanon is its straight up collusion at the Board of Directors level between the two companies, and no one can convince me otherwise.

 

[ToTTenTranz]

Configs are defined by the primary market they sit in.

Oh really?

What was the Steam Deck's Van Gogh defined for, with a first-of-its-kind CVML block inside? And the Switch 1's SoC?

AMD says everything with DPFP at half rate is HPC. Too bad!

I guess it's a good thing that doesn't apply to the Radeon VII which is the Gaming GPU SKU you brought up yourself.

AMD’s Radeon VII GPU Will Not Support Uncapped Double-Precision (FP64)​

Really stretching the definition here.

Nope. It's really just the definition.

Is H100 (with its single gfx-enabled TPC) a client GPU?

It would be if someone sold it as a client dGPU.

Can you create "semantics" thread instead?

You don't need semantics discussion for statements like these, that are very clearly wrong.

Dawg this ain't 2007 anymore, AMD is long over client gfx.

No one at AMD is interested in gaming.

And I'm honestly done with giving this a pass. It's just not productive and borderline trolling.

 

[adroc_thurston]

What was the Steam Deck's Van Gogh defined for, with a first-of-its-kind CVML block inside?

Surface Neo.
Lakefield was the same.

I guess it's a good thing that doesn't apply to the Radeon VII which is the Gaming GPU SKU you brought up yourself.

The die has native 1/2 DPFP. It's not a gaming part. Never was and never will be.

It would be if someone sold it as a client dGPU.

okay you're deluded

 

[reaperrr3]

But the way they're treating the software side is what convinces me of this "we don't care about dGPU" theory. I can't think of any reason why they refuse to bring FSR4 to older gens other than "we care so little about this market we're gonna fk over loyal customers lololol"

They refuse to bring FSR4 to older gens because
- fully debugging the INT version of it would be software work for cards already sold long ago, not making them any new money
- you're supposed to upgrade every gen, not use FSR to buy yourself time and skip their shiny new cards!!!

 

[ToTTenTranz]

Surface Neo.

What's Surface Neo? What SKU code is that? Can I buy one?
No? Then it's just another one of those cases where a chip originally designed for one thing ends up in another thing with a different purpose.

I don't understand the "we're pivoting away from dGPUs" move at precisely the point when AMD has their best dGPU arch relative to nvidia in the past 15 years.

You don't need to understand something they're not doing.

But the way they're treating the software side is what convinces me of this "we don't care about dGPU" theory. I can't think of any reason why they refuse to bring FSR4 to older gens other than "we care so little about this market we're gonna fk over loyal customers lololol"

FSR4 is 99% coming to RDNA3/3.5. It's the timings and messaging that are typical AMD's "never miss a chance to miss a chance" moto.

The die has native 1/2 DPFP. It's not a gaming part. Never was and never will be.

It was you who mentioned the Radeon VII name. Which AMD presented, marketed and sold as a gaming part.

They refuse to bring FSR4 to older gens because
- fully debugging the INT version of it would be software work for cards already sold long ago, not making them any new money
- you're supposed to upgrade every gen, not use FSR to buy yourself time and skip their shiny new cards!!!

AMD's latest product release with a RDNA3.5 part was yesterday.

 

[adroc_thurston]

Then it's just another one of those cases where a chip originally designed for one thing ends up in another thing with a different purpose.

It's the same purpose, VGH remained a tablet chip.
You don't like, have to be intentionally obtuse.
Reality will hit you hard anyway.

It was you who mentioned the Radeon VII name. Which AMD presented, marketed and sold as a gaming part.

Is Vega20 a gaming part.
Is H100 a gaming part?

 

[ToTTenTranz]

It's the same purpose, VGH remained a tablet chip.
You don't like, have to be intentionally obtuse.
Reality will hit you hard anyway.

There are two products using Van Gogh:

1 - Steam Deck (spanning 3 SKUs), a handheld gaming device.
2 - Magic Leap 2, a transparent XR headset.


List of tablets with Van Gogh:
- None

Cope harder.

Is Vega20 a gaming part.

Radeon VII is a gaming part that uses Vega 20.

 

[MrMPFR]

Can you create "semantics" thread instead?

Yes please. This BS has been going on for almost a week and no one is going to agree on anything. So please either agree to disagree or move discussion somewhere else.

You don't need semantics discussion for statements like these, that are very clearly wrong.

Depends on how you define it.

 

[adroc_thurston]

Steam Deck (spanning 3 SKUs), a handheld gaming device

That's a tablet.

Radeon VII is a gaming part that uses Vega 20.

Is H100 a gaming part?

 

[ToTTenTranz]

That's a tablet.

It's a gaming handheld. Not a tablet. Absolutely no one at AMD who supplies the SoC, Quanta who assembles it or Valve who designed and sells it, defines the Steam Deck as a tablet.

Steam Deck™

Steam Deck™ is the most powerful, full-featured gaming handheld in the world.

store.steampowered.com

Is H100 a gaming part?

Radeon VII is a gaming SKU, which you wrongly claimed it wasn't.

Depends on how you define it.

No, it does not.
Saying "no one at AMD is interested in gaming" doesn't depend on any kind of interpretation gymnastics, it's factually incorrect and it's only tiresome trolling that needs to be called out.

 

[adroc_thurston]

Absolutely no one at AMD who supplies the SoC, Quanta who assembles it or Valve who designed and sells it, defines the Steam Deck as a tablet.

It's a tablet.

Radeon VII is a gaming SKU, which you wrongly claimed it wasn't

Is H100 a gaming part? It has a single TPC for gfx, after all.

Saying "no one at AMD is interested in gaming" doesn't depend on any kind of interpretation gymnastics, it's factually incorrect and it's only tiresome trolling that needs to be called out.

It's so factually incorrect they have zero dedicated gaming dGFX parts now.
too bad! but you're the type to enjoy scraps methinks.

 

[Win2012R2]

It's a tablet.

And what is Switch 2 - a robot?

It's so factually incorrect they have zero dedicated gaming dGFX parts now.

And what is 9070 then?

 

[marees]

And what is 9070 then?

he is probably referring to the RDNA 5 line up where AT0, AT2, & AT4 have a tie up with microsoft to achieve scale

AT3 (medusa halo / 10060xt) is an outlier as there is no product tie up (except for the strix halo like dgx/rocm box but not sure how it is viable unless memory prices crash)