Discussion RDNA 5 / UDNA (CDNA Next) speculation

[marees]

Feel free to speculate

no such thing as UDNA.
There's RDNA5, and it has 3 parts and no halo.
Baby, token market presence with no ambition to take any heads whatsoever. Ergo, they're dead.

RDNA 5 potential features
👇

Shared L1 Has Far Reaching Implications

Global LDS patent Patent available here: https://patents.google.com/patent/US12001334B2 It's a radiating fabric paradigm. Does nearest neighbour searches first, then propagates out to other CUs further out. The patent is old so likely major tweaks since then (LDS instead of L0 etc...). But the ...

docs.google.com

Description of few patents:
👇

Page 32 - Discussion - RDNA 5 / UDNA (CDNA Next) speculation

Page 32 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

More patents
👇

Page 65 - Discussion - RDNA 5 / UDNA (CDNA Next) speculation

Page 65 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Possible configs

Page 28 - Discussion - RDNA 5 / UDNA (CDNA Next) speculation

Page 28 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Summary :— https://www.jonpeddie.com/news/amds-next-generation-gpu-architecture-udna-and-rdna-5/

 

[Saylick]

Rumored roadmap:


Emphasis on lack of halo SKU.

 

[marees]

Rumored roadmap:
View attachment 113792

Emphasis on lack of halo SKU.

Thanks got the source (below)

I find it very hard to digest that RDNA 5 will launch in 2027

Surely GPUs can't be that hard ???

https://twitter.com/x/status/1833491928739930330

 

[adroc_thurston]

Surely GPUs can't be that hard ???

8q cadence is pretty standard.

 

[SolidQ]

Emphasis on lack of halo SKU.

Still don't see nothing about any part just NAVI5
Let's wait in future about HALO happens/not happens

 

[Saylick]

Still don't see nothing about any part just NAVI5
Let's wait in future about HALO happens/not happens

There’s no box in the Ultra Enthusiast category for RDNA4 or 5, which, according to the slide, was previously occupied by N21 and N31.

 

[SolidQ]

was previously occupied by N21 and N31.

But navi 31 doesn't have Ultra Enthusiast, if 6900 was compete +- with 3090, 7900 not compete with 4090
and we need to know fake/not fake roadmap. Let's wait from CES roadmap

 

[adroc_thurston]

But navi 31 doesn't have Ultra Enthusiast, if 6900 was compete +- with 3090, 7900 not compete with 4090

that's because they've missed. It's the same swimlane.

 

[SolidQ]

that's because they've missed

let's wait new docs. at least if it's just Enthusiast, should compete with RTX 6080, it's fine then, but would be mega fail, if they not beat near 4+ year old 4090

 

[adroc_thurston]

should compete with RTX 6080

nope.

but would be mega fail, if they not beat near 4+ year old 4090

Well yeah. They've surrendered.
Just gave up.

 

[SolidQ]

They've surrendered.

That only 30% from XTX, they should do it. Feels like PS6 gonna pretty weak console, if they not gonna beat or queal 4090 with middle

 

[adroc_thurston]

Feels like PS6 gonna pretty weak console, if they not gonna beat or queal 4090 with middle

well yeah, shrinks ran out.
PS5/XSX were lucky to jump onto the "last good node shrink in CMOS history" train cab.

That only 30% from XTX, they should do it

maybe? irrelevant.

 

[Tup3x]

With roadmap like that I honestly don't see them gaining more market share. Halo SoCs are nice and all assuming that you can sell them - there's going to be more competition there... We'll see what happens.

 

[SolidQ]

With roadmap like that I honestly don't see them gaining more market share

let's wait official roadmap, i don't trust those "leaked" roadmap. For me this roadmap looks like photoshoped, too much distortion
Navi 22 different numbers "2", Navi 44 reading like NaYi 44

 

[ToTTenTranz]

All this can change at any time, especially if the AI bubble bursts, ARM chips find a way to run x86 code decently and AMD needs consumer dGPUs to make money again.


Never again is such a terrible statement to do in high tech.

 

[adroc_thurston]

and AMD needs consumer dGPUs to make money again.

They make EPYC.
Consumer dGPUs never made a penny for AMD either way.
'cept maybe RDNA2.

 

[SolidQ]

Consumer dGPUs never made a penny for AMD either way.

Rly? What is that? Driven By Radeon. meaning RDNA3

 

[marees]

8q cadence is pretty standard.

In general, I agree with the 2 year cadence

There are 3 reasons, why I am hopeful of a shorter cadence this time:

1. As it is AMD has no 4080 competitor. I find it untenable that there won't be a 4080 / 5070 ti competitor until 2027
2. There are already 3 products on 3nm: zen 5c server chip, instinct chips, strix halo. Should be easy to do a die shrink from 4nm to 3nm & also create a 5070 ti competitor for enthusiasts
3. Xbox next should be out soon

 

[adroc_thurston]

Rly? What is that? Driven By Radeon. meaning RDNA3

Oh yeah, RDNA3 too I guess.

why I am hopeful of a shorter cadence this time

Shorter cadence is only for things that make money. I.e. GPGPU.

As it is AMD has no 4080 competitor. I find it untenable that there won't be a 4080 / 5070 ti competitor until 2027

Not AMD's problem.

There are already 3 products on 3nm: zen 5c server chip, instinct chips, strix halo

Turin-D yes, MI350X is H2 next year and halo is niche.

Should be easy to do a die shrink from 4nm to 3nm & also create a 5070 ti competitor for enthusiasts

You're not getting a shrink either. It's gfx12 straight to gfx13. Pure IP focus, they're ignoring product-level opportunities completely.
Only gfx IP itself matters since it goes into APUs.

 

[Tigerick]

Preliminary RDNA5 Speculation

RDNA3	Die Size	VRAM	BW		RDNA4	Die Size	VRAM	BW		RDNA5 / Blackwell	Die Size	VRAM ?	BW ?
N31	529 mm2	24GB 384-bit GDDR6	960 GB/s							RDNA5-1		24GB 256-bit GDDR7	1152 GB/s
										RTX 5080	377 mm2	16GB 256-bit GDDR7	960 GB/s
										RTX 6080		36GB 384-bit GDDR7 ?	1440 GB/s ?
N32	346 mm2	16GB 256-bit GDDR6	620.8 GB/s		N48	?	16GB 256-bit GDDR6	640 GB/s		RDNA5-2		18GB 192-bit GDDR7	864 GB/s
										RTX 5070		12GB 192-bit GDDR7	672 GB/s
										RTX 6070
N33	204 mm2	8/16 128-bit GDDR6	288 GB/s		N44	?	8GB 128-bit GDDR6	320 GB/s		RDNA5-3		12GB 128-bit GDDR7	576 GB/s
										RTX 5060		8GB 128-bit GDDR7	448 GB/s
										RTX 6060

• Based on leaks, we should be expecting three monolithics dies of RDNA5. I am using Micron GDDR7 roadmap to speculate upcoming GDDR7 speed and capacity.
• I am always amazed by NV's GPU layout and die size especially compared to AMD's N31 chiplet design.
• It seems RDNA5 finally catch up to Blackwell design albeit two years later. By that time, NV should be launching RTX-6000 series, make you wonder what new features NV going to launch?
• Based on memory bandwidth, RDNA5-1 should be faster than RTX-5080 (And RTX-4090) provided AMD don't screw up the design...
• The 3GB memory die should be standard by end of 2026, that's why RDNA5-1 should come with 24GB GDDR7 as standard. That explains slightly higher tier than N48 in the slide above.
• Again, this is early speculation to jump start the discussion, feel free to provide your idea...

 

[branch_suggestion]

Based on leaks, we should be expecting three monolithics dies of RDNA5. I am using Micron GDDR7 roadmap to speculate upcoming GDDR7 speed and capacity.

Probably, but maybe they do something neat with the top part, though nothing that distracts for Instinct.

I am always amazed by NV's GPU layout and die size especially compared to AMD's N31 chiplet design.

RDNA3 missed on clocks, the point was to reduce BoM at a modest PPA hit, but the PPA was not as good as hoped.

It seems RDNA5 finally catch up to Blackwell design albeit two years later. By that time, NV should be launching RTX-6000 series, make you wonder what new features NV going to launch?

?

Based on memory bandwidth, RDNA5-1 should be faster than RTX-5080 provided AMD don't screw up the design...

Not a problem on N3P

The 3GB memory die should be standard by end of 2026, that's why RDNA5-1 should come with 24GB GDDR7 as standard. That explains slightly higher tier than N48 in the slide above.

Depends on if 256-bit is the end result.

Again, this is early speculation to jump start the discussion, feel free to provide your idea...

2 years on debating a <$1k generation in a dead market and a dead profession.
Sigh.

 

[Win2012R2]

I am going out on a limb here to speculate that RDNA5/UDNA1 will be very disappointing...

AMD needs new RADical leadership that will tolerate no failures - big or small...

 

[marees]

I am going out on a limb here to speculate that RDNA5/UDNA1 will be very disappointing...

AMD needs new RADical leadership that will tolerate no failures - big or small...

I have some hope that xbox next would be based on RDNA 5. So some good features might come from that partnership

 

[Win2012R2]

I have some hope that xbox next would be based on RDNA 5

Xbox is dead - PlayStation won, maybe it's a good thing and Microsoft will finally get serious at getting Windows up to console level when it comes to running games (properly supporting unified memory for PC games ports to avoid multiple copying of same data.

 

[adroc_thurston]

Xbox is dead

Nope. But they need to start making more actual games and actually try marketing them.