Discussion RDNA 5 / UDNA (CDNA Next) speculation

[ToTTenTranz]

View attachment 129600

Did you just do full circle between @Kepler_L2 's post in this very same forum and wccftech?

Glad that you're openly admitting to be trolling.
At least we now both know where we're standing.

 

[Josh128]

Did you just do full circle between @Kepler_L2 's post in this very same forum and wccftech?

Absolutely I did. I didnt see Keplers post here, so I went for it. Plus, the 'kek had some cool graphics. 🤣 🤣

 

[Magras00]

It's "unified" in the sense that CDNA is picking up some RDNA features and ISA (they're both supposed to be gfx13 right?)
They're not going to be selling MI configurations to gamers.

It seems like AMD want to go from Vega (GFX9.x) derived ISA for CDNA 1-4 and then to RDNA4 derived ISA for CDNA5 (GFX12.5) before the UDNA merger that probably comes sometime in 2027 (just a guess). GFX13 products first and prob then derived ISA and maybe foundational architecture for DC with some derived GFX13 version for CDNA6 or whatever it'll be called.

UDNA merger is not about architecture it's about merging ISA. Perhaps also about sharing a foundational architecture (SM structure, cache hierarchy...) between MI and RX like NVIDIA Blackwell DC vs consumer.
Jack Huynh talked a lot about not wanting to reset the matrix on the optimizations every time they do a memory hierarchy redesign. Unified cachemem hiearchy will for sure happen with UDNA, otherwise optimizations can't be ported between DC and consumer. AMD has been all over the place but if you look at NVIDIA they've been remarkably consistent: L2, L1, shared memory and some other SM level caches since basically Fermi (+15 years).
Everything else can and will vary wildly like NVIDIA to maximize PPA for each market.

Let's adress the RDNA4 derived ISA claim. GFX tag is about ISA and not architecture. It literally says "shader ISA" before GFX tag in TPU's GPU database. Architecturally RDNA 2 changed multiple things but besides some besides accommodating the new stuff (like RT) the core ISA is almost identical to RDNA1. As a result GFX ISA is almost the same.
RDNA1 dGPUs = GFX10.1
RDNA 2 dGPUS = GFX10.3

RDNA3 had significant ISA changes and major changes (eq. Polaris -> Vega) to architecture while RDN4 also had major changes to architecture and ISA. Based on C&C's LLVM analysis from last year it looks like the ISA changes are actually significantly larger than what was the case with RDNA3 but correct me if I'm wrong here.
RDNA3 dGPUs = GFX11.0
RDNA3.5 iGPUs = GFX11.5
RDNA4 dGPUs = GFX12.0

Now compare the Vegas with the CDNA's.
Vega 10 = GFX9
Vega 20 = GFX9
Vega iGPU = GFX9
CDNA = GFX9
CDNA2 = GFX9
CDNA3 = GFX9
CDNA4 = GFX9

And now for the big moment

CDNA5 = GFX12.5

Like CDNA's don't mean it's architecturally aligned with RDNA4 but simply that it has a RDNA4 derived ISA. Vega worst GCN ISA baggage just has to go at some point. CDNA 6 is prob when MI and RX merge into UDNA with the conditions explained earlier.

From what someone told me a while back AI/ML ops aside the ISA in RDNA is still mostly similar to GCN despite the significant µArch changes?

Seems like RDNA4 did a lot to change ISA as mentioned ^, but yeah in general RDNA didn't wipe the slate clean like GCN. It just adressed the weaknesses of GCN and built upon it. IDK about ISA, but the schedulign is nothing like GCN's antiquated scheme. There's a great slide from AMD in the C&C RDNA 3.5 compiler post showcasing the difference: https://chipsandcheese.com/p/amd-rdna-3-5s-llvm-changes

My fear with this is RDNA5 needing a complete driver overhaul that will massively hinder the GPUs' performance at launch, resulting yet again in those "Fine Wine" months / years where AMD loses all the marketshare because release-date performance is reallistically all that matters as most GPU reviews happen in those months.

This at best indicative but the RDNA5's local launcher patent shared by Kepler (not a leak) specifically mentioned how the scheduling system has a fall back option where it defaults to the old way. They won't wipe the slate clean without a fallback option like @soresu said they need PS4 and PS5 BWC. It's very important.
But they need to invest heavily in that driver team. Cannot afford another RDNA driver disaster.

 

[adroc_thurston]

NVIDIA Blackwell DC vs consumer.

Completely different L1, L2 and mem implementations.
And let's not even talk about TMEM.

Cannot afford another RDNA driver disaster.

There never was one?

 

[Magras00]

Was talking about the cache hierarchy not specifics. Again foundational architecture, maybe I should've been even more specific and called it a shared template. Yes these are very different between consumer and DC.

Here's the almost 1 year old quote from Jack Huynh in case anyone is interested:
"So, one of the things we want to do is ...we made some mistakes with the RDNA side; each time we change the memory hierarchy, the subsystem, it has to reset the matrix on the optimizations. I don't want to do that.

So, going forward, we’re thinking about not just RDNA 5, RDNA 6, RDNA 7, but UDNA 6 and UDNA 7. We plan the next three generations because once we get the optimizations, I don't want to have to change the memory hierarchy, and then we lose a lot of optimizations"
Link: https://www.tomshardware.com/pc-com...na-together-to-take-on-nvidias-cuda-ecosystem

TMEM roughly equivalent to a private VRF for Tensor cores is basically scratchpad for systolics (see #1,231) and split into four, one within each SIMD partition, not a cache. Don't see how this is relevant to the cache hiearchy discussion.

Remember this issue heavily discussed after the RDNA launch. Most coverage around late 2019 early 2020 IIRC. Really hurt AMD mindshare, if there was anything left by then.

 

[adroc_thurston]

TMEM roughly equivalent to a private VRF for Tensor cores and split into four, one within each SIMD partition, not a cache

It's neither cache, not a regfile.
It's basically a scratchpad for systolics.

Remember this issue heavily discussed after the RDNA launch. Most coverage around late 2019 early 2020 IIRC

That wasn't drivers.
Navi10 just shipped with a bucketload of h/w bugs.
Didn't impact AMD "mindshare" in any way, either. They had none.

 

[soresu]

Remember this issue heavily discussed after the RDNA launch. Most coverage around late 2019 early 2020 IIRC. Really hurt AMD mindshare, if there was anything left by then.

Mindshare for enthusiasts maybe, but mindshare in general has little to do with that.

Don't get caught on YT tech channel talk as being indicative of the general buying public that mostly buy whole, prebult systems, rather than DIY builds from separate components.

 

[adroc_thurston]

the general buying public that mostly buy whole, prebult systems

Oh noes, prebuilts are a relatively tiny market.
What moves AIC volume is DIY sales.

 

[soresu]

Oh noes, prebuilts are a relatively tiny market.

Wait even OEM prebuilts?

 

[adroc_thurston]

Wait even OEM prebuilts?

Those are even smaller (for gaming, anyway).
This isn't 1996 anymore.

 

[regen1]

 

[eek2121]

https://videocardz.com/newz/amd-rdna5-gpus-reportedly-get-transformer-codenames-alpha-trion-ultra-magnus-orion-pax

Can we not just have an awesome discussion without a bunch of money hungry morons making a post out of every single comment?

I'm against discord because it isn't accessible from search engines, but maybe we do need to move to discord...

 

[poke01]

https://videocardz.com/newz/amd-rdna5-gpus-reportedly-get-transformer-codenames-alpha-trion-ultra-magnus-orion-pax

Can we not just have an awesome discussion without a bunch of money hungry morons making a post out of every single comment?

I'm against discord because it isn't accessible from search engines, but maybe we do need to move to discord...

I think so too. Any leaks should be shown in the Anandtech discord.

Videocardz and WCctech should find their own sources. Freeloading critters

 

[poke01]

Or we can do what Famiboards does and lock leaks to Anandtech account only users

 

[Kepler_L2]

Or we can do what Famiboards does and lock leaks to Anandtech account only users

Or we can start a new business, OnlyLeaks.

 

[adroc_thurston]

I think so too. Any leaks should be shown in the Anandtech discord.

Videocardz and WCctech should find their own sources. Freeloading critters

Who cares it's a public forum.
Everything interesting is discussed far from here anyway.

 

[Saylick]

Or we can do what Famiboards does and lock leaks to Anandtech account only users

Well, they could just as easily create a lurker account…

 

[poke01]

Who cares it's a public forum.
Everything interesting is discussed far from here anyway.

It is public, yes. Anyway no point discussing, ethics in journalism died a long time ago

 

[adroc_thurston]

ethics in journalism died a long time ago

It never existed.

 

[GodisanAtheist]

https://videocardz.com/newz/amd-rdna5-gpus-reportedly-get-transformer-codenames-alpha-trion-ultra-magnus-orion-pax

Can we not just have an awesome discussion without a bunch of money hungry morons making a post out of every single comment?

I'm against discord because it isn't accessible from search engines, but maybe we do need to move to discord...

-When the Anandtech Forums finally get shutdown expect half a dozen rumor mills to close shop.

 

[igor_kavinski]

Suggestion: when sharing previously unknown information, preface with "This is NOT a leak and this is known fairly widely..." and then watch the leak-leechers squirm as they decide how to share the source that says specifically that it's not sharing a leak

 

[adroc_thurston]

Suggestion: when sharing previously unknown information, preface with "This is NOT a leak and this is known fairly widely..." and then watch the leak-leechers squirm as they decide how to share the source that says specifically that it's not sharing a leak

I say that all the time but clickbait farmers don't care.

 

[Saylick]

Suggestion: when sharing previously unknown information, preface with "This is NOT a leak and this is known fairly widely..." and then watch the leak-leechers squirm as they decide how to share the source that says specifically that it's not sharing a leak

That won't work. All they do is just flag what they are "reporting" as a rumor so that they can absolve themselves of all liability. Even when the rumor is later debunked, all they do (if they even do it) is update the article saying that the rumor has been debunked, but by that point they would've garnered the clicks and ad impressions already, and business operations goes on as normal. Imo, as I mentioned previously, the best way to counter them is to embed some form of personal insult that is difficult to remove (e.g. watermark) so that they risk getting mocked on the internet if they decide to source your info.

 

[soresu]

Or we can start a new business, OnlyLeaks.

Bruuuuuh that still sounds way too dirty 😂😆

 

[igor_kavinski]

It is being reported from one of the packaging partners that a *WCCFTECH leechers discouraged*mysterious CPU box is being *WCCFTECH leechers discouraged*manufactured. The source does not yet have the front of the box *WCCFTECH leechers discouraged* but according to them, the back of the CPU has exciting information with important keywords like "supercharge your games *WCCFTECH leechers discouraged*with the best *WCCFTECH leechers discouraged*Turbo Boost yet" and "AVX-512 helps to make even slower*WCCFTECH leechers discouraged* GPUs faster by removing critical *WCCFTECH leechers discouraged*bottlenecks exposed in the execution pipeline *WCCFTECH leechers discouraged*due to legacy code by introducing *WCCFTECH leechers discouraged*a new optimized codepath that offers better power efficiency and performance without stratospheric temperatures. Over 100 games*WCCFTECH leechers discouraged* supported at launch!".