Discussion RDNA 5 / UDNA (CDNA Next) speculation

[Tuna-Fish]

There's a "Medusa Premium" between Point and Halo now, right? Wouldn't that make for a decent handheld?

 

[MrMPFR]

Related to DGF/DMM, at 20:25, One of the AMD RT Fellow Architects is asked that can DGF and DMM be combined and he basically says yes but it will require architecture work to do so and the goal is to have a DGF base layer with DMM encoded on top of the DGF base layer.

Well that certainly explains why AMD has multiple DMM patents. Couldn't get my head around this because NVIDIA deprecated it but this makes sense. Oh and those patents list the former NVIDIA veterans Holger Gruen and David McAllister, that were poached by AMD a few years ago. The patents were also shared by Kepler ~2 months ago.

This suggests AMD's DMM related R&D efforts are as serious as their DGF efforts. Implementation of both in RDNA 5 very likely, although still not confirmed IIRC.

Edit: Found these patents relevant for the animated DMM paper shared by @basix:

US20250299447A1 - High-quality skinned object animations for displaced micro meshes - Google Patents

A technique for performing graphics processing is provided. A displaced micro-mesh is created that includes a plurality of coarse base triangles that define a set of coarse vertices each of which has an associated directional value. The displaced micro-mesh also includes a set of displacement...

patents.google.com

US20250104285A1 - Lossy geometry compression using interpolated normals for use in bvh building and rendering - Google Patents

Devices and methods for rendering objects using ray tracing are provided which include generating a low resolution version of a high resolution mesh representing objects in the scene, determining points on curved surfaces of curved surface patches defined for one of triangles and bi-linear...

patents.google.com

 

[basix]

Here's an animated DMM paper from AMD:

https://gpuopen.com/download/AnimatedDMMs.pdf

 

[marees]

There's a "Medusa Premium" between Point and Halo now, right? Wouldn't that make for a decent handheld?

Yes that is likely the z3e

Also it is very similar to a series S (if you consider Magnus the series X)
Medusa Premium = 24 CU AT4 (10050) Lpddr5x 128 bit/ lpddr6 192 bit
Magnus = 72 CU AT2 (10070) Gddr7 192 bit

 

[MrMPFR]

Here's an animated DMM paper from AMD:

https://gpuopen.com/download/AnimatedDMMs.pdf

TL;DR (not complete):

1. In mid 2024 AMD proposed a novel method for ray tracing of animated DMMs devoid of lighting artifacts.
2. Lossy geometry compression but since removes geometric artifacts if prior implementations the visual quality difference vs full mesh negligible .
3. Method described as "...interpolating (vertex-based) transformation matrices first and using the result to transform DMM vertex and normal data."
4. Compression ratio for Kermit the Frog asset is 20X. From ~10 MB to ~0.5 MB.
5. Simplification to base mesh reduces "...triangle count by a factor of 64, which amounts to a DMM of subdivision level 3." ie from 400K to 6K tris.
6. Performance penalty (RX 7900XT) is significant: "Due to higher algorithmic complexity and register pressure, our approach has ∼ 2.3−2.5× higher ray tracing costs than regular ray tracing of static DMMs."
7. Makes DMMs applicable to all sorts of micro-triangle geometry, overcoming limitations of prior art.
8. Base mesh can be built once since animations doesn't affect rebuilds.
9. Multiple ideas in the pipeline for reducing performance overhead and expanding upon the implementation.

#8: DGF doesn't have this functionality even with the animation update. While the implementation in the paper differs, the end result is similar to the cluster templates used in Animated Clusters, a part of Mega Geometry. RTX MG relies heavily on tessellation and subdivision surfaces, so not surprised that AMD is exploring DMMs.
This is pure speculation but I don't think the deprecation for DMMs gives the full picture since the GitHub documentation indicates that NVIDIA might still be using the HW 40 series introduced to accelerate some aspects of RTX MG, but it's possible that it's just standard tesselation.

Seeing Carsten Benthin being involved makes AMD's DMM push even more serious. He is one of the authors behind Intel's ray tracing of micro-polygon geometry (Nanite-esque) paper published in August 2023, nearly 1.5 years prior to RTX Mega Geometry albeit it is a less encompassing implementation. Benthin is another RTRT R&D heavy weight with decades of experience.

Edit: Bad wording fixed and thank you @soresu for the PDF link and oops I forgot. Intel link to paper + PDF download included now.

 

[adroc_thurston]

Seeing Carsten Benthin being involved in the makes DMM push even more serious. One of the authors behind Intel's ray tracing of Nanite-like geometry paper and another R&D heavy weight with decades of experience.

oh hey look you've actually done some linkedin-fu.

 

[MrMPFR]

oh hey look you've actually done some linkedin-fu.

Nope. Boring.
Much more interesting to go through research papers and patent filings.

See prev comments getting a lot of likes so before RDNA5 hype runs out of control let's temper expectations. For DMM and DGF in HW all we got rn is AMD's comment about DGF decompression in HW in February and Kepler confirming DGF for RDNA 5. Everything else is up in the air and SW =/= HW. The extent of the hardware DGF implementation in RDNA 5 remains unknown and any DMM patents shared by Kepler was a response to a question for what beyond Blackwell RT feature set could be. In order words besides HW DGF decompression none of it is confirmed as it stands.

 

[soresu]

Seeing Carsten Benthin being involved makes AMD's DMM push even more serious. He is one of the authors behind Intel's ray tracing of Nanite-like geometry paper published in 2023 and another R&D heavy weight with decades of experience.

Direct link to the Intel paper pdf as you can only request access from that site.

 

[Saylick]

Huge AT0 news

Free xcloud tier (ala GeForce Now)

https://twitter.com/x/status/1974132006838841848

Backing off Xbox consoles and pivoting Game Pass to cloud gaming.

 

[marees]

View attachment 131493

Backing off Xbox consoles and pivoting Game Pass to cloud gaming.

That's just a salty nobody

Check post by HeisenbergFX4 in same page

News - Microsoft revamps Xbox Game Pass plans and hikes Ultimate to $29.99 a month, Ultimate gets Ubisoft+ Classics and more

Xbox Game Pass, the most UNaffordable deal in gaming :messenger_tears_of_joy: Halo won't be back

www.neogaf.com

 

[Saylick]

That's just a salty nobody

Check post by HeisenbergFX4 in same page

News - Microsoft revamps Xbox Game Pass plans and hikes Ultimate to $29.99 a month, Ultimate gets Ubisoft+ Classics and more

Xbox Game Pass, the most UNaffordable deal in gaming :messenger_tears_of_joy: Halo won't be back

www.neogaf.com

This one?

“but my confidence is way lower than my 99% I have been saying”

 

[marees]

This one?
View attachment 131494
“but my confidence is way lower than my 99% I have been saying”

Note: *contracts with AMD*

 

[marees]

This one?
View attachment 131494
“but my confidence is way lower than my 99% I have been saying”

Latest
👇

it was our sources that confirmed the existence of the Xbox Series X|S, their specs, their price, in addition to reports on the cancelled cloud "Keystone" console, the shelved first-party Xbox handheld, and more recently, the Xbox Ally. Codenamed Kennan and comprised of Omni (Xbox Ally) and Horseman (Xbox Ally X), the Xbox Ally range is part of Microsoft's multi-year partnership with AMD, which will include tailor-made devices from first party.

I've asked multiple trusted sources, many of whom spoke to me on the above devices previously which turned out to be true, about what Xbox's present hardware posture is. For now, it remains unchanged.

Microsoft publicly confirmed that it's still committed to first-party Xbox hardware over the summer, with comments from both President Sarah Bond and AMD CEO Lisa Su on silicon that would power next-gen Xbox devices — both from third and first-party.


From what I've been told, the hardware plans are not even "up in the air," and remain firmly Xbox's path forward. Xbox Series X|S hardware production has not ceased, and new stock will go out to retailers at its usual cadence.

No, Xbox's next gen console hardware plans aren't cancelled — for now, but it's a problem that it was really easy to believe

Rumors are swirling that Xbox is going to kill its hardware ecosystem. It's not true ... for now, but it easily could've been true, given how nonsensically Xbox acts lately.

www.windowscentral.com

 

[adroc_thurston]

Latest
👇

Yeah like Xbox hardware plans haven't changed ever since they acquired Zenimax.

 

[maddie]

This one?
View attachment 131494
“but my confidence is way lower than my 99% I have been saying”

A division of a company can declare bankruptcy to void contracts? Sounds like a recipe for chaos?

 

[branch_suggestion]

AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs

News Highlights OpenAI to deploy 6 gigawatts of AMD GPUs based on a multi-year, multi-generation agreement Initial 1 gigawatt OpenAI deployment of…...

ir.amd.com

MI450 is quite good.

 

[jpiniero]

AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs

News Highlights OpenAI to deploy 6 gigawatts of AMD GPUs based on a multi-year, multi-generation agreement Initial 1 gigawatt OpenAI deployment of…...

ir.amd.com

MI450 is quite good.

I dunno, AMD's stock is up 25% premarket. It's like they do this because they know AMD stock will skyrocket. OpenAI buys a billion dollars worth of AMD GPUs, gets stock worth 25 billion for free.

 

[branch_suggestion]

I dunno, AMD's stock is up 25% premarket. It's like they do this because they know AMD stock will skyrocket. OpenAI buys a billion dollars worth of AMD GPUs, gets stock worth 25 billion for free.

This was in the works for a good while and OAI won't see any ROI until the first tranche is delivered and if the stock is valuable enough.
MI450 good, roadmap good.

 

[basix]

MI450 good, roadmap good.

This is exactly what I was thinking. OpenAI gambles on AMD stock pricing going up. Basically GPUs for free. But that won't work if the MI450 and the further roadmap are not promising enough.

 

[jpiniero]

This is exactly what I was thinking. OpenAI gambles on AMD stock pricing going up. Basically GPUs for free. But that won't work if the MI450 and the further roadmap are not promising enough.

Unless they sell before the AI bubble bursts.

It's lke Sam has found an infinite money glitch. That's not even including if they did any... umm.... insider trading... before the announcement was made.

 

[CakeMonster]

AMD associating with that grifter makes me more skeptical of their prospects. But the market seems hellbent on riding this trend, so I guess the actors act accordingly because stock price.

 

[basix]

160 Mio. shares of that deal are now worth 32bn USD. If AMDs market cap rises to the sky like Nvidia has (3x and therefore "only" ~30% of Nvidias current market cap), those shares are worth 100bn USD.

Or in other words: 2 Mio. GPUs 50k$ each for basically free
- 2 Mio GPUs with 2kW each sound adequate for a 6GW deployment

 

[Win2012R2]

But that won't work if the MI450 and the further roadmap are not promising enough.

They will know what MI500 will look like.

Good deal for AMD, that should put billions into improving stupidly named ROCm framework, perhaps CUDA's moat will finally be bridged...

 

[adroc_thurston]

AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs

News Highlights OpenAI to deploy 6 gigawatts of AMD GPUs based on a multi-year, multi-generation agreement Initial 1 gigawatt OpenAI deployment of…...

ir.amd.com

MI450 is quite good.

 

[Z O X]

AMD associating with that grifter makes me more skeptical of their prospects. But the market seems hellbent on riding this trend, so I guess the actors act accordingly because stock price.

A little OT, but what is stopping them from lending compute power, similarly to AWS?
With Palatnir, and who knows what else, income is guaranteed even if LLMs for general public stagnate.
Line go up ...