Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

Man I have been dying to make this one for a while now.

First rumours for RDNA3 are here so new thread time!

Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[H T C]

Yes and no.

Nvidia's efforts at re-use / dual use have mostly been that though. AD102 is a bit too new, but look at GA102:

15 cards, 7 are gaming, 8 are not.

So while fusing off things, or locking drivers is one thing, the silicon is design to do both.

Gaming cards have things disabled that professional cards do not, right?

However, i have NO IDEA if any of those disabled things are RT related or not.

 

[biostud]

Better AI, I'm skeptical. In a shooter it's trivial.

I haven't played it myself but I remember project F.E.A.R being renowned for good AI, where the AI worked with the bots in groups, for different team strategies.
So I think there are vast improvements for AI also in shooters if you look at is a group AI.

 

[KompuKare]

Gaming cards have things disabled that professional cards do not, right?

Sure, and sometime it might even be due to yield but that's probably very rare.

My point was about the cost of designing the chips: while GA100, DA100 etc. is purely a compute cards, a lot of Nvidia's "Gaming" have silicon which has to do other workloads and it seems rather than spending money designing two dies they have found a way to re-use those part. Potentially wasting silicon on gaming cards, but hey their margins are high enough anyhow.

 

[Mopetar]

Red Faction was fun for the destructible environments. I remember creating a lot of tunnels instead of actually playing the game properly a lot of the time. The multiplayer maps were a lot of fun as well and really tapped into that vibe and scene from the Matrix where at the end you had this shot up and destroyed lobby.

It certainly does have limits as I remember hitting the point where if you made a long enough tunnel it would either have to undo the changes made at the start (I think the first game did it this way) or you'd eventually hit a limit where it wouldn't let you destroy any more of the environment (pretty sure the second game handled it this way).

It was great fun though and you could do some fun stuff like making your own little cave to hide in and snipe with the rail gun. Make the geometry of it properly and anyone trying to storm it had to cross your field of view and could easily be shot before they could shoot you.

 

[Saylick]

Red Faction was fun for the destructible environments. I remember creating a lot of tunnels instead of actually playing the game properly a lot of the time. The multiplayer maps were a lot of fun as well and really tapped into that vibe and scene from the Matrix where at the end you had this shot up and destroyed lobby.

It certainly does have limits as I remember hitting the point where if you made a long enough tunnel it would either have to undo the changes made at the start (I think the first game did it this way) or you'd eventually hit a limit where it wouldn't let you destroy any more of the environment (pretty sure the second game handled it this way).

It was great fun though and you could do some fun stuff like making your own little cave to hide in and snipe with the rail gun. Make the geometry of it properly and anyone trying to storm it had to cross your field of view and could easily be shot before they could shoot you.

Haha, I remember playing Red Faction 2 and doing something like what you described as well. Just hide in a hole and snipe people with the X-ray vision scope of the rail gun.

Anyways, I've said this a few times on these forums but you can put me squarely in the "RT is nice to have, but not needed" camp. Until a game implements RT in a way that is vital and crucial to the gameplay, RT will just be visual eye candy like when Nvidia pushed tessellation super hard with Fermi. AMD were behind on tessellation by a factor of 2 or more and Crysis 2 happened to be the game that was the poster child for advocating Fermi (remember those concrete road barriers?). Now, instead of tessellation we have RT, and instead of Crysis 2 we have Cyberpunk 2077 with Psycho RT settings.

 

[Timmah!]

But professional workloads are USUALLY done with professional cards: NOT with gaming cards, which the 4090 clearly is, no?

Not neccesarily. Why would you buy Quadro for cold hard cash, if Geforce can do the same and is cheaper? Not everyone doing pro work is a big corporation or university with money no object attitude, so lets just buy the most expensive stuff there is, because it says "professional" on the box, regardless of the fact whether its actually needed. Then again thanks god for them, cause if they started money-pinching and buy Geforces instead of Quadros, i could see Nvidia quickly do away with Geforce support of this non-gaming stuff.

 

[Timmah!]

Gaming cards have things disabled that professional cards do not, right?

However, i have NO IDEA if any of those disabled things are RT related or not.

The tensor stuff is usually artificially limited, as its used for some of that AI stuff and they want you to buy Quadros for that purpose. Limiting RT would limit gaming performance, so Geforces are not limited in there. Aside of not having full chip, like 4090 being cut to 128 SMs out of 144, when Quadro will be no doubt full fat 144SM chip.

 

[Yosar]

Why perceived? It was clearly stated 2300Mhz for the SE, which mostly contributes to the power and performance. Frontend has way less switching capacity, so it could be clocked higher without tradeoffs. As for "sync", afaik amd uses asymmetric IFOPs for the short routes since zen2, where any uncore clock domain can be "decoupled"

Perceived because maybe due to reasons of architecture they can not be clocked higher for now. If something can not be clocked higher you can not say it's clocked low. It's not the same. Because where is that limit you consider good enough? 3 GHz? Why not 5 GHz? After all CPU chips are clocked that high.
Are nVidia chips clocked low because they are nowhere near 5 GHz? And after Zen 4 we all know TSMC 5nm is capable to be clocked that high.

All those opinions of low clock are based on 'leaks' (and we know where we should put them now) and the fact that RDNA2 was clocked higher in weaker models (because 6900XT was actually clocked lower than 7900 XTX is now).
Only RDNA2 against Navi 31 was like almost totally different architecture. And it makes difference.

It's interesting that in no marketing/teasing material AMD ever said RDNA3 will be clocked much higher than RDNA2. Contrary to Zen 4 where almost any tease was about chips being clocked much higher than Zen 3.
RDNA3 was all about architecture.

 

[Kepler_L2]

Perceived because maybe due to reasons of architecture they can not be clocked higher for now. If something can not be clocked higher you can not say it's clocked low. It's not the same. Because where is that limit you consider good enough? 3 GHz? Why not 5 GHz? After all CPU chips are clocked that high.
Are nVidia chips clocked low because they are nowhere near 5 GHz? And after Zen 4 we all know TSMC 5nm is capable to be clocked that high.

All those opinions of low clock are based on 'leaks' (and we know where we should put them now) and the fact that RDNA2 was clocked higher in weaker models (because 6900XT was actually clocked lower than 7900 XTX is now).
Only RDNA2 against Navi 31 was like almost totally different architecture. And it makes difference.

It's interesting that in no marketing/teasing material AMD ever said RDNA3 will be clocked much higher than RDNA2. Contrary to Zen 4 where almost any tease was about chips being clocked much higher than Zen 3.
RDNA3 was all about architecture.

"It (RDNA 3) is also our first gaming GPU architecture that will leverage the enhanced 5nm process and an advanced chip packaging technology. And another innovation includes a architected compute units with enhanced ray-tracing capabilities and an optimized graphics pipeline with even faster clock speeds and improved power efficiency." - David Wang @ FAD 2022

"We’ve driven the frequency up, and that is something unique to AMD. Our GPU frequencies are 2.5 GHz plus now, which is hitting levels not before achieved. It’s not that the process technology is that much faster, but we’ve systematically gone through the design, re-architected the critical paths at a low level, the things that get in the way of high frequency, and done that in a power-efficient way.

Frequency tends to have a reputation of resulting in high power. But in reality, if it’s done right, and we just re-architect the paths to reduce the levels of logic required, without adding a bunch of huge gates and extra pipe stages and such, we can get the work done faster. If you know what drives power consumption in silicon processors, it’s voltage. That’s a quadratic effect on power" - Sam Naffziger @ Venture Beat

So I don't think it's fair to say that clock speed increase was only expected because of leaks.

 

[DisEnchantment]

The more I read about this clock thing, sounds like another jebait operation from Herkelman. I bet it must be too tempting for him to let this slide.
The gap between AD102 and AD103 left an opening for AMD. I wonder if N32 is competitive with AD104, if they can maintain this same segmentation.

7900XT/XTX puts the 4080 in a weird position which the jebait team of Herkelman & Azor keeps going on and on. Quite boring for the average tech guy.
If AMD can afford to play the market share card, then great for them. They are likely not losing money selling this chip with all the surplus 5nm/6nm in the post pandemic era. And does not matter if they use a lot more silicon to beat a smaller chip. Nobody bothered if the 4090 is a lot bigger chip with 20 billion more transistors and beating smaller N31 chip.

I find it mildly annoying with the hush hush attitude with AMD. Product launched, but another NDA for clocks really?
I am inclined to believe this clock topic is another jebait operation.

 

[moinmoin]

Nvidia has been a good teacher to AMD, the jebaiting is not letting up.

I personally think this is mainly a cover up for the botched spec which I do think is lower than targeted. But AMD also learned well to make the most of what they actually have in hand at a given time. So they do just that.

 

[jpiniero]

with all the surplus 5nm/6nm in the post pandemic era.

If anything, with the market rejecting the original N3 in favor of N4, N5/N4 is going to continue to be very busy.

 

[Dribble]

Apparently you never played Red Faction where almost the entire game was destructible. It most certainly can be done. Also, Battlefield has had destructible environments for over a decade, going way back to Bad Company 2.

Not realistic, those have very set limits on what could blow up and how it could blow up. e.g. in BC2 you can shoot a house and it probably has about 30 different ways it could be destroyed depending on where you hit it so in fact you can store all the destroyed state for a house in a 10 bit number. That's very different to obeying physics and having a near infinite number of outcomes.

 

[Glo.]

The more I read about this clock thing, sounds like another jebait operation from Herkelman. I bet it must be too tempting for him to let this slide.
The gap between AD102 and AD103 left an opening for AMD. I wonder if N32 is competitive with AD104, if they can maintain this same segmentation.

7900XT/XTX puts the 4080 in a weird position which the jebait team of Herkelman & Azor keeps going on and on. Quite boring for the average tech guy.
If AMD can afford to play the market share card, then great for them. They are likely not losing money selling this chip with all the surplus 5nm/6nm in the post pandemic era. And does not matter if they use a lot more silicon to beat a smaller chip. Nobody bothered if the 4090 is a lot bigger chip with 20 billion more transistors and beating smaller N31 chip.

I find it mildly annoying with the hush hush attitude with AMD. Product launched, but another NDA for clocks really?
I am inclined to believe this clock topic is another jebait operation.

To some degree, it might be, but it appears that GPUs are not able to clock to 3 GHz with sane power.

And If I understand this architecture correctly, clock speeds are the most important for NGG paths, and is making front end much more important in the grand scheme of things. The higher the clock speeds - the more you cull. The higher the culling of geometry, the higher the performance you get.

This basically explains why there are two different clock speeds for ALUs and for front end.

Also, the 2.8 GHz wall(at the moment) that everyone talks about - its shader clocks. Front end clocks - should be different, higher.

So all of those leaks, and all of what people are touting, that those GPUs might clock to 3, or higher GHz - it may be only about front end clocks. Not the shader clocks.

 

[Joe NYC]

If anything, with the market rejecting the original N3 in favor of N4, N5/N4 is going to continue to be very busy.

There was some report saying the utilization on N5 is likely to drop to 90% (from 100%) in coming quarters.

 

[Kaluan]

I'd have a lot to add or reply to but I currently have a nasty cold and I can't focus worth much.

Anyway, I'll mention my biggest observation... Has anyone noticed that (at least to my knowledge, correct me if I'm wrong) that AMD never mentioned 7900XT/7900XTX dies by their (code)name in the slides and official press coverage? All they say about the 300mm+5-6x37mm SKUs is "Navi3x", never specifically the widely believed "Navi31". Any explanation on that? Thanks.

 

[Stuka87]

Not realistic, those have very set limits on what could blow up and how it could blow up. e.g. in BC2 you can shoot a house and it probably has about 30 different ways it could be destroyed depending on where you hit it so in fact you can store all the destroyed state for a house in a 10 bit number. That's very different to obeying physics and having a near infinite number of outcomes.

Later battlefields had more pre-programmed destruction. But in BC2, you could level pretty much every house. It was very common for there to be no buildings left by the end of a battle.
NOTE: Dice went away from the full destruction because they didn't like how the maps turned into big open areas, as it ruined any type of dynamic game play. So they removed a lot of destruction when BF3 came out.

As for Red Faction, nothing was pre-programmed. It used physics. You could literally drill tunnels all over the place. That was its claim to fame. Full blown physics driven destruction.

EDIT: A post from one of the devs on Red Faction Physics:

https://www.reddit.com/r/truegaming/comments/639btc/_/dfsqzcv

 

[uzzi38]

I'd have a lot to add or reply to but I currently have a nasty cold and I can't focus worth much.

Anyway, I'll mention my biggest observation... Has anyone noticed that (at least to my knowledge, correct me if I'm wrong) that AMD never mentioned 7900XT/7900XTX dies by their (code)name in the slides and official press coverage? All they say about the 300mm+5-6x37mm SKUs is "Navi3x", never specifically the widely believed "Navi31". Any explanation on that? Thanks.

There's nothing to say about it. If you're getting your hopes up for an RDNA3 SKU that's larger than Navi31 with some twisted logic of "maybe this isn't N31 and there's something bigger": don't.

This is without a doubt N31 and the largest RDNA3 die that will come to market.

 

[biostud]

This is without a doubt N31 and the largest RDNA3 die that will come to market.

The largest single GCD die RDNA3

 

[uzzi38]

The largest single GCD die RDNA3

...no dual GCD products will come to market either

 

[Timorous]

If there is an issue that means it clocks lower than desired at a given voltage and it is fixable the best I can see is a refresh mid next year with 7950XT and 7950XTX adding maybe 20% more performance at the top end.

It feels like AMD were gunning for just over 4090 raster perf at a more sensible price point but got close enough to not want to bother delaying knowing their refresh products will be that little bit better instead.

 

[Kaluan]

There's nothing to say about it. If you're getting your hopes up for an RDNA3 SKU that's larger than Navi31 with some twisted logic of "maybe this isn't N31 and there's something bigger": don't.

This is without a doubt N31 and the largest RDNA3 die that will come to market.

I disagree with your straw manning assumption.

There's plenty to say, if N31 indeed has hardware bugs, then we may as well see N31a, N31b or N31 XT1, XT2 whatever.

This has nothing to do with bigger dies or whatever.

I feel like a lot of people are spazing out in whatever the opposite of hype mode is. Overcompensating and all. Chill out people.

 

[uzzi38]

I disagree with your straw manning assumption.

There's plenty to say, if N31 indeed has hardware bugs, then we may as well see N31a, N31b or N31 XT1, XT2 whatever.

This has nothing to do with bigger dies or whatever.

I feel like a lot of people are spazing out in whatever the opposite of hype mode is. Overcompensating and all. Chill out people.

A refresh would still be called Navi31.

 

[biostud]

...no dual GCD products will come to market either

I know, just want to fuel the hype train. Toot toot.

 

[Kaluan]

A refresh would still be called Navi31.

I supppse, still I think it was a interesting tidbit to notice.