Support NeoGAF

[Proelite]

It's the same size as RDNA3.
I scaled, then trace over each block.

RDNA3.5 is even wilder.
The WGP is constantly changing.
Like Kepler said, it'll be bigger.

I'm looking it up and they are narrower than in your illustration. Even your RBE is wider than necessary. I think you can comfortably do 5/6WGP per row instead of 4.

Navi 48 XL, RDNA 4.

Stryx point RDNA3.5

 

[thicc_girls_are_teh_best]

Moore law is dead, any console more powerful will cost more and will not get cheaper overtime

More expensive means less public willing or capable of buy which in turn means more risk to publishers to launch games only on the console

So the question remains, if they launch a console called PS6 and all or almost all the games continue to be made and launched for and on the ps5 is this truly a PS6 or is just a PS5 Pro 2?

That's less a problem of Moore's Law, more one of Dennard Scaling these days.

That's a good point you've raised, I'm curious what the narrative and discourse will be going into "next-gen", and surrounding these consoles. Sony will likely. and based on past history go down an efficient route, like they did with PS5's "fast and narrow design", relatively less CU's but a faster clocked GPU or Xbox's slower but wider approach, more CU's but clocked relatively slower.

If we assume Magnus is Xbox's next chip, then it's likely they'll be sticking with the same design philosophy, but this time ensuring there's a clear advantage in performance compared to the competing PS6. Personally I welcome it, but I'm curious how the market will react to a product will likely could be priced $1000 +

What might also complicate things is the ML technology, PS5 Pro using PSSR, and will soon incorporate into it FSR 4, I'm wondering if Microsoft will also adopt FSR 4 and how this may complicate things considering Project Amethyst. I'm wondering if Microsoft may develop their own ML based upscaling technology.

It really just comes down to the value proposition MS can offer. Exclusives 100% won't be a part of that, since they release their games everywhere now. Price won't be it, at least not with any Xbox device using this Magnus APU.

At the end of the day it has to still deliver the convenience & usability of an Xbox console environment, but some of the more obvious openness & modularity of PC & Windows, in a way where they don't conflict with each other.

Considering that PS6 is releasing in 2027+, I don't see how they can guarantee a substantial power difference when PS6 will be on n2p and the Series X2 will be on n3p. It's a 40% density gap. That's not their goal because it's not going to be a realistic goal. Their goal is to be the best gaming PC one can get in a console like hardware / form factor (i.e. APU + soldiered GDDR7 main memory) that can rival RTX 5080 / 4090 performance in 2026. By 2028/2029, they'll have a successor that will trounce the PS6 handily.

Hypothetical PS6 in 2027 will have
8-10 core Zen 6
24GB GDDR7 on 256 bit bus (896GB/s bandwidth)
64-128mb Infinity Cache
68/72 Cus at 3.5 ghz (up to ~61 teraflops)
whichever forward UDNA 2/3 features Cerny decides to incorporate due to Project Amethyst.
~300mm2 die size
$599-$699 (Most likely $699)
such a PS6 will likely perform as well if not better with optimized console SKUs of games.

Yeah those are pretty realistic PS6 specs, and the thing with MS's approach is they can justify 3-year iterations across APU packages if they want, so a high-end device in 2029 outperforming PS6 would be expected.

But that type of device is also going to cost a lot more, for obvious reasons.

Hello

Why would you assume the Xbox consoles won't have optimized SKUs?

SIE's SDK and APIs have generally been better than Microsoft's for over a decade now. This gen it got so bad with Xbox that "the toolz" became a meme.

Basically, PlayStation APIs are more tuned and focused on game development, while the DX12U APIs are more focused on providing a huge gamut of solutions for general 3D pipelines that can include (but aren't limited to) gaming. However, the degree of choice with implementations you can roll with in DX12U actually can be time-wasting for engines if they need to plug through a crap ton of things to find the one solution that works well with them.

That type of redundancy isn't really a thing with PlayStation APIs, AFAIK, so for game dev they tend to be better. Plus SIE have groups like XDEV, which grew from the ICE teams, which were specifically made to help come up with solutions for game dev on PlayStation platforms like PS3 and onward.

That's disappointing NGL. Sony and MS have always being within one year after volume production of a node since Xb1/PS4 gen.

I'm guessing the AI boom is affecting availability of advanced nodes.

Well, someone on IconERA pointed out the Rapidus fab in Japan having recently developed a 2nm chip, and from what I read that fab is being developed to help companies (including Sony, who have a big investment in its development) to be less reliant on TSMC for wafers.

It's a good call, and if the fab can get large enough scale in the next few years there's a decent chance SIE & Sony will use its 2nm for PS6. But it's nothing locked in stone ATM, and could very well not happen. Tho, it's just nice to see another promising fab come up that could become a viable alternative, after Global Foundries basically fell behind (and even Samsung, for the smaller nodes, more recently).

 

[paolo11]

How strong is this magnus cpu? 4090 level?

 

[Landr300]

That's less a problem of Moore's Law, more one of Dennard Scaling these days.

the cost stopped failing at 28nm a has been increasing since

https://en.eeworld.com.cn/news/manufacture/eic660660.html

its literally over in the economic side

if they choose to launch a new console on the market they can call it ps6 how many times they want but as soon as every game continues to launch in the ps5 the console will be a ps5 pro 2 in everything but the name

 

[wafflecioccy]

I don't think 24gb RAM is enough for a proper generational leap with PS6.

Consider the OS needs resources and such, and there are plenty of ideas for games that need a ton more RAM like for physics.

 

[Physiognomonics]

I don't think 24gb RAM is enough for a proper generational leap with PS6.

Consider the OS needs resources and such, and there are plenty of ideas for games that need a ton more RAM like for physics.

Could be enough if developers properly use PS5 fast I/O. They'll do it more if next Xbox is irrelevant to them.

But 3rd party developers really have to stop using CPU to stream in and decompress data on PS5 like in so many games still.

 

[Loxus]

I'm looking it up and they are narrower than in your illustration. Even your RBE is wider than necessary. I think you can comfortably do 5/6WGP per row instead of 4.

Navi 48 XL, RDNA 4.

Stryx point RDNA3.5

Tried that already.

It makes the GCD bigger than the SoC.
For it to work, the GCD has to be on 2nm, while the SoC is on 3nm.

That may be why he only listed the node for the SoC. The GCD is probably on a different node.

 

[wafflecioccy]

Could be enough if developers properly use PS5 fast I/O.

Which will never happen because exclusives aren't really a thing anymore.

 

[KeplerL2]

Tried that already.

It makes the GCD bigger than the SoC.
For it to work, the GCD has to be on 2nm, while the SoC is on 3nm.

That may be why he only listed the node for the SoC. The GCD is probably on a different node.

That was my thought too, but since the video came out I went through a bunch of internal AMD documents and managed to find a mention of Magnus among a bunch of other N3P CPU/GPU/APU products in development so I think N3P is more likely.

 

[Radical_3d]

well, next gen will need serious horsepower as that generation will be the raytracing generation.

current gen is the beginnings of raytracing, and next gen will be when it will be matured. similar to how gen5 was the first steps into 3D, while gen6 finally was actually good at 3D.

AI upscalers only help so much when it comes to raytracing.
if you play Doom TDA with a very low internal resolution, you'll start seeing very obvious artifacts from the drastically reduced ray count.
this doesn't happen with Doom Eternal as it doesn't rely on RT GI.
so at the same internal resolution Eternal will look better/more stable than Dark Ages, which means you can use more aggressive reconstruction without introducing to many issues.

ray reconstruction can help with with a bit, but can also introduce its own artifacts.


so both manufacturers need to make sure they have enough horsepower that raytracing, and possibly pathtracing run well and look clean.
you don't want to be the system that has constant boiling artifacts or excessive ghosting compared to the other one.

Ray reconstruction. That's all it'll be about in next gen. We've already see it on the SW2. A weak ass hardware looking reasonably well. For the average Joe it'll be indistinguishable. If you watch a DF video with direct comparison sure. Here are more upscaling artefacts, etc. But in general to appeal to the higher end user with a more expensive machine is going to be suicidal. I bet most will be content with the "more affordable" PS5 Pro. What a time to be alive, to write that last sentence.

 

[Toots]

It is definitively a Xbox apu, they choose Madusa (not Medusa), the woman who single handedly revived WWE lady's division, as a symbolic counter to the greatest playstation fanboy :

the Hulkster himself

 

[Proelite]

That was my thought too, but since the video came out I went through a bunch of internal AMD documents and managed to find a mention of Magnus among a bunch of other N3P CPU/GPU/APU products in development so I think N3P is more likely.

Have you shared all you know about upcoming consoles?

 

[Clear]

The problem MS has is that they'll need to refresh all their Azure racks to support XCloud.

One key point I think that kinda got lost in the weeds was that from the start the back-end was designed to run clusters of Series S level hardware. So a massive jump in power at the base level is going to incur an equivalently huge increase in both installation and running costs for their ever so important streaming ambitions.

Unless they are pivoting away from their Xbox Anywhere approach back-end provision/running-cost has to be a larger area of concern than the contents of their consumer units.

 

[Proelite]

Tried that already.

It makes the GCD bigger than the SoC.
For it to work, the GCD has to be on 2nm, while the SoC is on 3nm.

That may be why he only listed the node for the SoC. The GCD is probably on a different node.

Something like the below doesn't work? I made the WGP as narrow as the ones in RDNA3.5 and 4. Also the RBE should be much narrower because it's not increasing in transistor count much if at all.

K KeplerL2

 

[Physiognomonics]

Something like the below doesn't work? I made the WGP as narrow as the ones in RDNA3.5 and 4. Also the RBE should be much narrower because it's not increasing in transistor count much if at all.

K KeplerL2

We can expect UDNA CUs to be bigger than RDNA4 because of improved RT. I expect 4 Shader engines in order to put more CUs. XSX was already unbalanced in that area (vs PS5 and all others AMD RDNA2 GPUs) because it was also designed for compute servers. Could be the same here.

6 Shader Engines for a console doesn't make any sense. SEs take too much precious space.

 

[StereoVsn]

One major problem with that. Wafer prices have increased drastically.
A 300mm^2 chip in N6 costs around 70-80$
But in N3P it will cost around 150-170$
And in N2P it will cost around 260-290$

Either the console price has to increase a lot, or Sony will have to pick a smaller chip and clock it higher.
The later is what Nvidia and AMD have been doing with their GPUs.

And Sony could stay on older process node as well. These consoles aren't mobile phones so staying on N3P or even older is still valid.

And hell, Switch 2 seems to be doing ok on the older Samsung node even.

 

[Loxus]

Something like the below doesn't work? I made the WGP as narrow as the ones in RDNA3.5 and 4. Also the RBE should be much narrower because it's not increasing in transistor count much if at all.

K KeplerL2

If we use that shape, the GCD becomes taller than the SoC.

 

[Proelite]

If we use that shap, the GCD becomes taller than the SoC.

? All I did is make your WGPs and RBEs narrower. Is there something I am missing?

 

[Loxus]

? All I did is make your WGPs and RBEs narrower. Is there something I am missing?

You made the WGP narrower without taking into account it's area.

Making it narrower would also make it taller than in your edit because of it's area.

Keep in mind you're removing transistors.

 

[Proelite]

You made the WGP narrower without taking into account it's area.

Making it narrower would also make it taller than in your edit because of it's area.

Keep in mind you're removing transistors.

Do we have a number for area of each WGP in RDNA5? What is the area you used for the WGP?

Each WGP in the Navi 48 XL on n4 should be around 3.6mm2.