Next Gen GPU architecture GCN, RDNA, Navi 10/20 (PS5 Navi Hybrid, Xbox Navi Pure) *spawn*

On that subject, let’s look at the RTX line.

Is Turing basically a Pascal with added RTX and tensor cores? (Real question)

If the compute part of Turing is for all intents and purposes a Pascal, then what’s stopping AMD from calling a GCN with bolted on RT (and other) tech a ‘new architecture’?

 

It's not.
There's a whack ton more features it supports over pascal that we only see at the highest end of feature support which likely required some changes to the architecture to support it.

the full white paper is here:
https://www.nvidia.com/content/dam/...ure/NVIDIA-Turing-Architecture-Whitepaper.pdf

But if we ignore what needed to be changed to support RT Cores and Tensor Cores, and any type of data centre processing stuff, the support for GDDR6 or changes to nvlink etc.

We still need to look at things like variable rate shading, mesh shading, multi-viewport rendering, texture space shading, then new compressions methods etc. I don't see that being supported outside of the latest architectures.

Each as generation brought forward new architecture to support newer feature sets, that doesn't necessarily mean it largely won't look the same.
I think the move towards generic compute has sort of 'locked' architectures in a specific way. If we move away from generic compute perhaps we'll get absolutely whole new architectures, but at this point in time, we should be expecting evolutions in the architecture to support new features.

 

It's one of the reasons I specifically made sure I didn't mention Xbox.

Although I still expect Scarlett to be Navi based, about a month ago I put forward a hypothetical situation how PS5 could be navi and Scarlett could be next gen, in this case pure RDNA.
Didn't want people to think I was just trying to use this to support that line of thinking.

 

Yea me 2. I wanted to tread really carefully around this topic because people tend to get sensitive over power wars.
Ignoring the power talk, it does look like Sony made a real effort at trying to get this console launched in 2019 with everything they wanted.
I think they still have a chance at doing something here much earlier than MS. Even a 6 month lead is massive.

OG Xbox One was late to the game, which is why we saw them struggle so hard this time around. I would be curious to see how MS would respond with having 6 months less to work with.

 

I didn't join that conversation as I think there's a big grey area.
Although the changes could be big enough that no one would question if it's the same uarch and not just the next version.

I've not got anything to add to that part of the discussion to be honest. I'm just going by face value.

 

The core architecture is about how the GPU processes work and keeps the ALU's occupied. Terascale'sVLIW required shaders to fit a particular ideal model to maximise utilisation. GCN changed that to be more flexible. Over time, it's added schedulers and whatnot. Meanwhile, nVidia has different workloads using different functional blocks, which would have to be bolted on to a GCN design. A clean-slate design might find a better way to map workloads onto ALUs, reducing silicon complexity and improving relative performance and integrating latest features like RT and ML acceleration into the shader design, or into the chip architecture overall in the case of them having discrete blocks.

We definitely shouldn't be counting GPU architecture as 'cemented' at this point. RDNA could be a completely new design, or a GCN remaster, or a GCN remake. Incidentally, rDNA is a thing in biochem, where you inject new DNA into existing; the implication there being RDNA is GCN with stuff inserted rather than a new species altogether.

If 6 months is enough to gain a significant architectural advantage, Sony should wait. If it's not, 6 months waiting nets MS nothing. Is it realistic that AMD had two offerings on the table, Navi 10 in 2019/2020 and a significantly better Navi 20 six months later, and Sony chose the former for the sake of six months lead sales? My guess would be that AMD had general timelines, and perhaps Sony committed to a 2019 release and got what they have, and MS decided to go with the next-gen architecture and will get what they get when it's ready, which could be anything 12+ months after Sony. The gamble being that though later, and potentially late (18+ months), MS will have a significantly stronger/more economical offering and something that's better future-proofed for their cloud ideas.

Because all things being equal, MS launching the same time as Sony without any particular advantages will just see them make the same sorts of numbers as OXB and XBO. They either need notably more power for the same price, or some sidewise USP. MS winning console market is highly unlikely, and they're all about moving people over to the XBox cross-platform ecosystem. XBox Game Pass has released on PC, for example. Design choices would thus favour their long-term goals rather than fitting the traditional console generation cycling, which would mean skipping a unified launch to have a stronger machine, conceding the console sales war to Sony and instead focussing on the best devices to slowly grow their ecosystem of more and more subscribers. Let Sony have their 100+ million consoles; MS will take the 25 million ultra-core high-end console gamers with their monster machine releasing 2021, and another 10 million from their low-end StreamBox, and all of them will start to subscribe to Live services along with others such that, by the time we're looking at PS6 coming out, MS has 100 million subscribers playing games on PC and mobile and some consoles and that income isn't going anywhere when PS6 appears.

 

Fair enough, I'm not going to commit to a poor choice of words, but we have yet to see this though right?
I mean, has nvidia's architectures largely changed (have we had a clean slate architecture) since fermi/kepler?
And this also coincides well from when GCN hit the market.

My main argument is whether or not the evolutions in architecture matter, and I think they do, at least enough to be named something else.

 

whoa big edit addition

Yea, that's more or less a reasonable take of the situation. It's possible Sony committed to hit an October 2019 launch as a direct response to the late arrival of X1X in 2017. And then possibly blind sided by the industry support towards Ray Tracing as announced in March 2018, which could have been (still been) a non-issue if it never gain's real support.

Yea this makes reasonable sense. I don't know what the plan is with streaming and the current generation xbox's today. If the consoles of today, 1S, SAD, 1X all support games of next generation because of streaming, then their base is already solidified and you're right in that it makes sense to move slower and push their program forward and take losses than to rush.

 

Not comparing MS to Sony I'll give my thoughts if this article is true why they could've gone for navi lite instead of navi pure.

Timescales

• Navi pure is due 2020 and is a "new uarch" , so chances of delay may have been seen as more likely.
• Navi lite would be out earlier being a good pipe cleaner for the arch.
• Sony may have been aiming for 2019. (who knows, but worth adding)

PS4 BC

• As the generation has gone on, it has become arguably more important. Navi lite being hybrid would make it easier to customize adding changes to support PS4.
• Navi pure may have been a magnitude of difficulty harder to support BC.
• Cost associated to implementing it in lite compared to pure.

Performance

• Navi lite may not have a big performance deficit, if at all. Or one that would be able to be overcome with a slightly bigger or faster apu.
• Able to achieve performance characteristics that they require.
• Risks associated with new uarch

This all may be worth it due to the whole package. Navi pure may have an initial higher cost, then add everything else on top. For a system that may not net a lot more performance?
Possibly better to invest in SS storage technology, if it was about balancing the system.

 

Sony said a long time ago that their next Gen consoles would not be held back by older console generations, suggesting they were targeting a disruptive new cycle(no backcompat guaranteed), I guess at this time they were planning going 100% to that supposedly pure "RDNA" Navi 20, but, i suppose as well, they were not expecting the ~100 million consoles target to be met, neither that backcompat would be that highly demanded and regarded by players as it was the case somewhat success of XBX.
If Navi 20 is that much different than current GCN they probably got back to the drawing board to make a hybrid arch to allow 100% out of the box back compat at least with PS4, afterall this would allow a seemless transition for their 100 million users.
I think the big problem with Sony is that they bet everything on PS4 since there were rumours at the time that this would be the last console generation and i heard that they believed on it as well, so was last resource short term strategy to see what they could achieve, while MS on the other hand had PC with them, so no major worries and more time to develop a more long term strategy.
AMD on the other hand has no interest to completely lose it's gains from the console space on the PC space, games before Playstation 4 generation, on PC were heavily optimized for Nvidia+Intel archs, and now we see quite a different scenario, so maybe Navi 10 would allow a better transition on the PC space, as well, without losing performance on titles that take real advantage of the actual GCN major points.
If the rumours are true and Navi has a more "Nvidia like" behaviour, so let's say a more balanced arch, without a major performance jump, Navi could undeperform in some titles compared to vega or polaris, which would be really odd on reviews.
And there is absolute no guarantee that it would perform on par in games more optimized for Nvidia anyway, so it would look in reviews some what like a regression not having a big win in "AMD titles" neither on "Nvidia ones."
In case of Nvidia from kepler to turing they acctually rebalanced their arch to be more GCN like, as we can see, in many games, so hence why i think this generation Nvidia jump was quite smaller than the one's before seems async compute was really a maxwell achiles heel and from now on will be heavily used in games.
Sorry for the english, and sorry for the extension on another topics, but i think they are heavily interlinked, and look too focused on one aspect make things looks too loose.

 

You made my text look from an neanthertal, you basically summed everything i want to say.

 

A lot of potential scenarios based on an unsubstantiated theory about a possible outcome that's unlikely.

The technical thread has become the baseless rumour thread and the baseless rumour thread has become a name of next gen console thread

 

Maxwell introduced some magic into Kepler.

 

The current going theory of PS5 going with an obsolete arch based on the unproven rumor that Sony was targeting a 2019 release its so silly.
It doesn't make any sense for Sony to gimp the PS5 GPU for BC, they'd never make the Wiiu mistake, PS5 would have enough raw power to just brute force past any incompatibility with a software layer. If BC was really an issue, it'd make more sense for Sony to invest a couple million on a software layer to assist hw emulation rather than cripple the GPU, that would be incredibly shortsighted, the CPU is a much bigger change anyways.

It would be in AMD's best interests to make sure PS5 uses the "pure" RDNA arch to guarantee another gen of games optimized for their GPUs. Its no coincide Lisa said RDNA would be Radeon graphics engine for the next decade while also commenting Cerny was going to revolutionize console gaming for the next decade.

Sony set a precedent this gen of implementing the latest (or even future) GPU features on their consoles.

It doesn't make sense to use a small 48CU chip on 7nm where they could fit a 80CU GPU on a 350mm2 APU die. You get much better perf/watt by going with more CUs clocked lower than less CUs clocked higher.
Realistically 64CUs (56 enabled-8 disabled) its the lowest PS5 GPU will go, leaving room for potential dedicated rt hw, audio chip and other silicon.

 

It wouldn't be going with an obsolete architecture, but going with the best available when they launch. Effectively the same as the G70 in PS3 when the notably better G80 launched close to (due to PS3 delays). If RDNA isn't ready, and only elements of it are, then a hybrid tech would be the only option for a console now.

Of course, a console now seems very unlikely given Sony aren't going to E3 this year. All of 4/5 months from release, and yet PS5 is still shrouded in mystery and isn't going to be announced?

 

I just read your previous post, you seem to agree Sony wouldn't go with an obsolete just to launch 6 month earlier.
The PS3/nvidia situation is nowhere near the same, we could argue over rumors of Sony asking nvidia for a gpu at the last minute (beggars can't be choosers) or we could go over what we know: Nvidia isn't interested in low profit console margins they wont do a semi custom let alone custom design, they'll offer whatever off the shelf part they got laying around. 8800GTX wasn't a good fit for a console considering its die size and thermals and nvidia wasn't going to design one just for the PS3 much less if it was a last minute call.

A console now isn't just unlikely, its already confirmed by Cerny no PS5 in 2019
What seems silly to me is this theory is based on unstained rumors and goes against Sony best interests and their history with AMD.

I mean if we gonna make theories based on rumors we could make all sorts of wild claims like MS going with Vega to double purpose Xbox APUs on their server blades

 

Wait, what? What did I miss?

 

There seems to be a lot of misinterpretation of what I was saying, or I'm misinterpreting you.
Why would navi 10/lite be gimped if the performance outcome is the same?
In regards to Navi pure we have no idea how that would translate based on the article I used as basis, also AMD old roadmap. I was just thinking through why it may not be navi 20 as has seemed to have been the ongoing theory:

• How big is navi pure, does it scale well to console characteristics?
• Would navi pure give worse performance in a console compared to Navi lite?
• TDP
• Would it have been ready in time.
• Given the amount of work that may be required to implement BC in software on a totally different uarch, would possibly also need earlier access.

You can debate how important BC is, but I would say the importance has only grown over the years.

There's many reasons not to go with the newest uarch. Navi lite would still be new and customised tech. Navi lite could be used in pc space for years and not be replaced in that segment by navi pure.

Obviously navi pure could check all the boxes for a console, but my point is there can be many reasons why it may not have been chosen. And none include PS5 ending up being gimped.

Luckily / hopefully will hear proper details soon.

All we have is rumors until details come out.
Just taking the rumors and having a discussion based on them.
You've said a lot like it goes against Sonys interest, yet I haven't seen why you think that apart from thrinking navi lite is gimped?

 

Think Sony has said PS5 isn't due this year.

But navi was due before this year, so maybe it impacted on Sony, maybe not and PS5 was always going to be 2020.

Sony has been very aggressive this gen and that's been nice.

 

It PS5 and Xbox have raytracing hardware, seeing the current Navi size, I think console will be chiplet based SOC not APU monolithic And Microsoft will probably have three models, Lockhart, Ananconda and Anaconda cloud(same SOC with HBM2 memory), it makes even more sense.

Sony will have Navi + 8 cores Zen 2, + maybe hardware raytracing and a 3d audio core. It seems difficult to have all this in a monotlithic APU.