Next-gen PS5 and next Xbox speculation launch thread |OT5| - It's in RDNA

[Pheonix]

I have a guess for one of the areas of differentiation in the PS5 and Xbox Scarlett silicon for next generation that could end up being talked about in Digital Foundry videos and such. Both sides have made mention of 8K and 120 fps support, which is somewhat checking the box for HDMI 2.1 compliance. However, to even provide any options for such things and arguably even still dealing with 4K resolutions and 60 fps on many games that are demanding, some type of "checkerboarding" or similar approach is going to be in play.

• PS5 - Digital Foundry and others have praised the approach to checkerboarding on the PS4 Pro, and I suspect that they will have some custom silicon (or "secret sauce" if you will) that provides support to an improved version of this intended to deal with resolutions all the way up to 8K as necessary.
• Xbox Scarlett - Notice that Navi does not include hardware support for Variable Rate Shading (VRS) which was disappointing to many, and VRS is definitely another approach to deal with the issue of scaling to different resolutions like with checkerboarding. Microsoft has been working deeply on VRS with DirectX (https://devblogs.microsoft.com/directx/variable-rate-shading-a-scalpel-in-a-world-of-sledgehammers/), and I suspect that their custom silicon will support VRS.

I am not going to predict which hypothetical solution may end up being better, but this does seem to be a possible area of differentiation given no Navi support for VRS.

Click to shrink...

8K right along with 120fps is just a buzzword. Completely meaningless one. They might as well just have said we will have HDMI 2.1. But saying 8K and 120fps just sounds better. Those consoles will struggle to hit 4K 60fps and we will see something very similar to this gen with 1080p where they will both run @4K, 30fps for most games and run at 4K, 60fps for a handful of games. No matter what, there will always remain that tradeoff of how much better a game can look if it were at 30fps as opposed to it being at 60fps. I hope I am wrong on the 4K,60fps thing but I doubt that.

I do however expect all 30fps games to have two gfx modes; Resolution and frame rate. With one focusing on 4K,30fps at High settings and the other focusing on 4K.cb,60fps with medium to high settings respectively. Should become the norm if the hardware for CB is baked-into the base models so that way it becomes a part of the dev pipeline as opposed to something that as to be made specifically for one platform (eg. PS4pro).

 

[Deleted member 12635]

110% 8k chat is because it's part of the 2.1 HDMI spec. Absolutely zero hardware silicon/work will be added to these machines in an effort to make possible.

Click to shrink...

110% --> over-utilized buzz word bingo!

The reason I am resistant to that is because people are using it to go all revisionist history on their 12-14TF predictions.

Click to shrink...

This indeed!


-------------------------------------

⁉ Thoughts ⁉

With up to 4 times the effective throughput you could argue that the moment developers will use Wave32 the average Performance Per Clock will improve again assuming that the numbers AMD provided were based on games still using Wave64.

 

[TUAXK]

I think project scarlett will definitely become navi 20. as it reads on the amd page

"" This processor builds upon the significant innovation of the AMD Ryzen™ "Zen 2" CPU core and a "Navi" GPU based on next-generation Radeon™ RDNA gaming architecture including hardware-accelerated raytracing. ""

* Based on next-gen Radeon. *

The film from Nvidia is not quite right. Somebody did the effort and calculated everything. with the power consumption
@ 18:20 ~

 

[bcatwilly]

8K right along with 120fps is just a buzzword. Completely meaningless one. They might as well just have said we will have HDMI 2.1. But saying 8K and 120fps just sounds better. Those consoles will struggle to hit 4K 60fps and we will see something very similar to this gen with 1080p where they will both run @4K, 30fps for most games and run at 4K, 60fps for a handful of games. No matter what, there will always remain that tradeoff of how much better a game can look if it were at 30fps as opposed to it being at 60fps. I hope I am wrong on the 4K,60fps thing but I doubt that.

I do however expect all 30fps games to have two gfx modes; Resolution and frame rate. With one focusing on 4K,30fps at High settings and the other focusing on 4K.cb,60fps with medium to high settings respectively. Should become the norm if the hardware for CB is baked-into the base models so that way it becomes a part of the dev pipeline as opposed to something that as to be made specifically for one platform (eg. PS4pro).

Click to shrink...

Man, you are probably right but I was hoping for 60 fps to be the next gen game standard. I will say that it is interesting that Sony chose to make the SSD the star of their Wired reveal, and Jason from Kotaku even noted that he had heard that "zero loading times" was expected to be a major marketing line by Sony for the PS5. And then you have Xbox where Phil Spencer has repeatedly talked about frame rates and overall system performance balance being the major Scarlett design goals every time he has talked about their next generation hardware. Of course the CPUs themselves will by default provide a nice boost to the situation, but at least going by things publicly so far I tend to think that Microsoft may put more of an emphasis on frame rates for their first party games at least (maybe mandate at least the types of modes you are citing?).

 

[BreakAtmo]

Same - they should focus on 60 fps instead with these new CPUs and throw as much GPU in there as they can for that 60.

Click to shrink...

God, if only. The CPU jump should absolutely allow them to go to 60fps while also seeing huge gains in simulation complexity and whatnot. I'm just worried that the GPU power that will also be needed might go straight to the visuals, even if they do use reconstruction techniques.

Designing all games to run at reconstructed 2160p60 would just be the best, IMO. And do 120fps for older games that can manage it.

 

[Straffaren666]

⁉ Thoughts ⁉

With up to 4 times the effective throughput you could argue that the moment developers will use Wave32 the average Performance Per Clock will improve again assuming that the numbers AMD provided were based on games still using Wave64.

Click to shrink...

For most use cases Wave64 will probably still be the most efficient mode for games, since there seems to be dependency stalls in Wave32 mode. If I understand it correctly, the major benefit of using Wave32 over Wave64 is the ability to run twice as many concurrent waves in Wave32 mode since it's using half the amount of resources. That's beneficial when running shaders with high register pressure (>64 VGPRs), but that's not the typical use case for games.

 

[BreakAtmo]

I was just playing Xenoverse 2, and I was thinking of another small way that SSD loading could improve a game. Whenever I'm playing a quest and a character transforms, it's done by pausing the action, cancelling all moves that have been made (frequently screwing me out of some Ki or Stamina) and going into a cutscene where they transform at a predetermined place. When that's done , they've been teleported to that place and I have to fly over to them.

Now, I'm assuming that this is a result of having to load in the new character model with new abilities in a predictable way. Does that mean that a next-gen exclusive Xenoverse 3 could replace all this annoying rigmarole with just having characters transform into their other forms and gain new abilities entirely in real time with no pause to the action?

 

[Putty]

Man, you are probably right but I was hoping for 60 fps to be the next gen game standard. I will say that it is interesting that Sony chose to make the SSD the star of their Wired reveal, and Jason from Kotaku even noted that he had heard that "zero loading times" was expected to be a major marketing line by Sony for the PS5. And then you have Xbox where Phil Spencer has repeatedly talked about frame rates and overall system performance balance being the major Scarlett design goals every time he has talked about their next generation hardware. Of course the CPUs themselves will by default provide a nice boost to the situation, but at least going by things publicly so far I tend to think that Microsoft may put more of an emphasis on frame rates for their first party games at least (maybe mandate at least the types of modes you are citing?).

Click to shrink...

Honestly, I applaud your excitement but think you're overthinking the CPU marketing.

 

[bcatwilly]

Honestly, I applaud your excitement but think you're overthinking the CPU marketing.

Click to shrink...

Fair enough, as everyone in this thread is probably guilty of overthinking things since we don't have all of the information yet :) But I still stand by the fact that these companies have certain overall design goals and focus when they set about doing these things even though the base is pretty much the same with AMD, and so far at least Xbox has really focused on the frame rate performance and such when discussing things (first mentioned by Phil Spencer at last E3 even).

 

[Deleted member 12635]

The main take-away of the performance of RDNA vs GCN (red rectangle)

 

[Miniature Kaiju]

For most use cases Wave64 will probably still be the most efficient mode for games, since there seems to be dependency stalls in Wave32 mode. If I understand it correctly, the major benefit of using Wave32 over Wave64 is the ability to run twice as many concurrent waves in Wave32 mode since it's using half the amount of resources. That's beneficial when running shaders with high register pressure (>64 VGPRs), but that's not the typical use case for games.

Click to shrink...

From the slides, Wave64 has potentially worse dependency stalls, actually.

 

[Straffaren666]

From the slides, Wave64 has potentially worse dependency stalls, actually.

Click to shrink...

Which slide is that?

 

[Deleted member 12635]

Which slide is that?

Click to shrink...

Slide 14.
I am of the same opinion as Miniature Kaiju

 

[Anthony Hopkins]

Next gens will have the rx twelfty hundred

 

[Anthony Hopkins]

My existing spec guess is 10-12TF GCN. What would that be in RDNA?

Click to shrink...

7 - 9 I think.

 

[Hey Please]

Previously AMD called a Shader Engine a whole slice of the GPU.

There was one Geometry Processor and Rasterizer, then the Shader Array consisting of a certain number of CUs (up to 16 under GCN) and the Rendering Backends (Up to 4 Render Backends/16 ROPs).
One Issue about the Shader Engine design for GCN was that with more CUs it was harder for the Shader Engine to feed them all if the workload was too short and the CUs would compute it faster than the SE could feed them.

The diagram for Navi10 shows actually no logical connection as to why it would be two Shader Engines.
In comparison to earlier it would be 4 Shader Engines (the blocks in red).

Under Navi 10 one slice seems to be 1 Prim Unit, 1 Rasterizer, a certain number of CUs (5 WGPs/10 CUs under Navi10), 4 Render Backends (16 ROPs) and 128KB of L1$ (at least for Navi10).
The Wave Launch Rate wasn't specified but AMD probably made sure that this wouldn't be an issue.

Click to shrink...

Thank you for this. So from what I understand the cluster comprising of the following is considered to be 1 WGP?:

• 2 CU (which share the following)
  • Vector Registers
  • Stream Processors
  • Schedulers
  • Scalar Units
  • Scalar Registers
  • Scalar Data Cache
  • Shader Instruction Cache
  • TMU
  • Texture Filter Units
  • Local Data Share (which presumably is one of the L0 or L1 cache?)

 

[Straffaren666]

The main take-away of the performance of RDNA vs GCN (red rectangle)

Click to shrink...

Slide 14.
I am of the same opinion as Miniature Kaiju

Click to shrink...

I agree with you, based on that slide it doesn't seem to be any advantage to execute in Wave64 mode. In Wave32 mode there are four cycles of dependency stalls and in Wave64 there are two cycles of dependency stalls. But to make a fair comparison we have to compare two concurrent Wave32 waves and one Wave64 wave. In that case the scheduler will be able to issue a salu instruction on the other Wave32 wave, eliminating the first salu stall and two valu instructions in the three cycle valu stall, resulting in only one valu cycle stall in Wave32 mode and two in Wave64 mode.

 

[isahn]

Current devkits likely have Vega 10/20 in them which can easily be at 12-14 Tflops. Navi in the retail units won't though.

Click to shrink...

indeed. From a dev point of view the API and in general the development environment of a new platform it's the most important thing to get an early hands on. And I'm sure that a Zen2+Vega PC can support the Scarlett/PS5 dev-env just fine.

Q: Impossible, hw ray-tracing and alien SSD?!
A: demoted to a software backend the one and PCI3 SSD for the other!

 

[M.Bluth]

⁉ Thoughts ⁉

With up to 4 times the effective throughput you could argue that the moment developers will use Wave32 the average Performance Per Clock will improve again assuming that the numbers AMD provided were based on games still using Wave64.

Click to shrink...

That was something I meant to ask about... For PC games, could we see a significant improvement for Radeon cards going forward if devs implement this?
If so, it'd be great for AMD in closing the gap between them and Nvidia in gaming performance.

 

[Straffaren666]

The main take-away of the performance of RDNA vs GCN (red rectangle)

Click to shrink...

For most use cases the throughput will be the same though. In this case, the per work-item IPC is only meaningful when there isn't enough concurrent waves to hide the latency. Generally, GCN will complete four times the work in four times the time, compared to RDNA.

 

[Locuza]

Thank you for this. So from what I understand the cluster comprising of the following is considered to be 1 WGP?:

• 2 CU (which share the following)
  • Vector Registers
  • Stream Processors
  • Schedulers
  • Scalar Units
  • Scalar Registers
  • Scalar Data Cache
  • Shader Instruction Cache
  • TMU
  • Texture Filter Units
  • Local Data Share (which presumably is one of the L0 or L1 cache?)

Click to shrink...

Consisting is a better term than "share" because certain logic like the Schedulers, Scalar Units, TMUs and the L0$ are not shared between two CUs but work either only per CU or SIMD32.

The Local Data Share is not a cache but scratchpad memory which is manually managed by the application, for example with Compute Shaders under DX11/12.
It's a seperate memory structure under GCN and RDNA:

Nvidia calls it Shared Memory which from an implementation standpoint changed multiple times from Fermi to Turing.
Pascal had it as a seperate structure being 96KB large (for 128 vALUs) and Turing has a unified memory structure which can be configured as 64KB Shared Memory + 32KB L1$ (the default mode for games) or as 32KB Shared Memory and 64KB L1$ (for 64 vALUs).

 

[DriftingSpirit]

At least I could keep up with the RAM talk. Now yall getting into stuff I can't handle. 😭

 

[Straffaren666]

That was something I meant to ask about... For PC games, could we see a significant improvement for Radeon cards going forward if devs implement this?
If so, it'd be great for AMD in closing the gap between them and Nvidia in gaming performance.

Click to shrink...

No, that IPC comparison is grossly misleading and will only occur in the most pathological use cases.

 

[orava]

How many tf's are rtx2080s?

Click to shrink...

~10 TFLOPS (FP32 float), but this does not tell that much about the actual performance difference. A RTX 2080 is generally faster when compared to a Radeon VII for example, which is almost a 14 TFLOPS gpu.

 

[dgrdsv]

Impossible, hw ray-tracing and alien SSD?!

Click to shrink...

SSD isn't the issue really, you can most certainly get similar throughput on PCIE 3.0 with a RAID or something.
RT can be ignored for the launch titles.

 

[Lady Gaia]

AMD won't have RT cores that are as big as AMD's.

Click to shrink...

I have an inside source that tells me otherwise. ;-)

 

[isahn]

SSD isn't the issue really, you can most certainly get similar throughput on PCIE 3.0 with a RAID or something.
RT can be ignored for the launch titles.

Click to shrink...

mh, I think you misunderstood my sarcastic Q/A post-scriptum to my post.
TL;DR I completely agree with you

 

[dgrdsv]

With up to 4 times the effective throughput you could argue that the moment developers will use Wave32 the average Performance Per Clock will improve again assuming that the numbers AMD provided were based on games still using Wave64.

Click to shrink...

The effective throughput of the CU isn't changing from GCN really. What's changing is the latency of instruction issue and the execution pipeline depth.
It was 4 16-wide SIMDs running each one wave over 4 clocks so you'd get 4x64=256 results in 4 clocks.
Now it's 2 32-wide SIMDs running each one wave over 1 clock so you get 2x32=64 results in 1 clock or 64x4=256 results in 4 clocks.
The fact that you can issue new instructions each clock and you get wave processed in one clock means a lot for pipeline stalls - the result of that can already be seen in 5700XT presumably beating Vega64 despite the latter having ~40% more processing power.

 

[Imitation Of Life]

Enhancements to checkerboard rendering and other reconstruction and upscaling techniques are important outside of 8K. I'd rather they not waste resources on native 4K either.

Click to shrink...

Reconstructed 4K and 60fps is ideal IMO.

 

[Gamer17]

Reconstructed 4K and 60fps is ideal IMO.

Click to shrink...

That's what's most probably gonna happen .or dynamic 4k 60

 

[formasymphonic]

can we get a couple of thread marks on these slides so that they are easier find / reference later?

 

[dgrdsv]

Considering that the new generation is starting at about twice the GPU power of mid-gen upgrades of the previous generation, I really doubt that we'll see many 60 fps titles.
People keep thinking for some reason that faster CPU is the only thing you need to get to 60.

 

[Miniature Kaiju]

Considering that the new generation is starting at about twice the GPU power of mid-gen upgrades of the previous generation, I really doubt that we'll see many 60 fps titles.
People keep thinking for some reason that faster CPU is the only thing you need to get to 60.

Click to shrink...

+1

We'll see 60 fps games, we'll possibly see more 60 than the current gen, but I don't think those will be anywhere near as ubiquitous as people seem to think (and definitely won't be an "across the board" thing).

 

[DriftingSpirit]

I just wouldn't go in expecting 60fps across the board either way. Some devs just don't do 60fps games regardless. There is more going into that decision than hardware.

 

[Pheonix]

Man, you are probably right but I was hoping for 60 fps to be the next gen game standard. I will say that it is interesting that Sony chose to make the SSD the star of their Wired reveal, and Jason from Kotaku even noted that he had heard that "zero loading times" was expected to be a major marketing line by Sony for the PS5. And then you have Xbox where Phil Spencer has repeatedly talked about frame rates and overall system performance balance being the major Scarlett design goals every time he has talked about their next generation hardware. Of course the CPUs themselves will by default provide a nice boost to the situation, but at least going by things publicly so far I tend to think that Microsoft may put more of an emphasis on frame rates for their first party games at least (maybe mandate at least the types of modes you are citing?).

Click to shrink...

Shouldn't be a problem for them if they are really pushing to continue supporting the XB1 or their platform free thingy.

 

[Fafalada]

• PS5 - Digital Foundry and others have praised the approach to checkerboarding on the PS4 Pro, and I suspect that they will have some custom silicon (or "secret sauce" if you will) that provides support to an improved version of this intended to deal with resolutions all the way up to 8K as necessary.

Click to shrink...

To be fair - Pro has other custom features that are related to resolution manipulation that haven't been widely talked about - part because they are usually associated with VR (not at all the only thing they are useful for), part because anything that requires highly custom code-path on a tiny install base of the premium consoles just doesn't get a lot of use, period.
But VRS is IMO orthogonal to all of this - it's not an either or scenario, having control over shading granularity is relevant to more than just output resolution IMO (speaking hypothetically here, as any serious real research work here is still in its infancy).

 

[Rylen]

the other issue is the xt has a 220 tdp, not useable for a apu. the 5700 is, but thats 8tfps. even if you remove board power, my gues is still around 200tdp for the xt.

Click to shrink...

It's 220w TBP, 180w TDP

 

[Miniature Kaiju]

By this time next year AMD will have a 5700 replacement that will get basically the same perf at 120W or less TDP. Mark my words.

 

[anexanhume]

It's 220w TBP, 180w TDP

Click to shrink...

Worth noting that part of the reason Nvidia is so much more efficient here is clocks in the 1600 MHz range for 2060 and 2070. Unfortunately, AMD has to clock Navi faster to keep up with these cards due to their architectural advantages. We can bet consoles will likely be 1600 MHz or below (my guess is 1400-1500) and that they'll sit at a happier place on the efficiency curve. They'll have to offset that with more CUs if they want to equal the above parts, though.

 

[Barnabas Reynolds]

None of us in this thread thought we'd see clocks that high until the Gonzalo leak. These parts have an impeccable track record of ending up in consoles. And frankly, there's several people in this thread just as knowledgeable on this subject as Richard, if not moreso.

Click to shrink...

That leak, which may or may not be related to PS5, and may or may not be representative of final, sustained clocks doesn't hold much water, imo, when the fact of AMD and Navi's capabilities in that area are in full view now. Everyone that is putting their faith in that leak is also expecting consoles to reach near-parity with nigh-on equivalent dGPUs, even though the former will be going into final production only months after the latter launches.

Not just Gonzalo but the leaks about PS5 devkits running even around 1850 Mhz not just 1800 Mhz.

Click to shrink...

What leaks?

 

[anexanhume]

That leak, which may or may not be related to PS5, and may or may not be representative of final, sustained clocks doesn't hold much water, imo, when the fact of AMD and Navi's capabilities in that area are in full view now. Everyone that is putting their faith in that leak is also expecting consoles to reach near-parity with nigh-on equivalent dGPUs, even though the former will be going into final production only months after the latter launches.

Click to shrink...

Provisional clocks in engineering samples go up in final products, not down.

The leak is almost assuredly linked to a next gen console. AMD has no other known semi-custom projects, and these chips following this numbering scheme have invariably ended up in consoles. Skepticism is always healthy, but dismissing a source of information because it doesn't agree with your personal beliefs just puts your bias on display.

Even if it reached those clocks and it was a Navi 10 die, it wouldn't hit the 10TF threshold many are hoping for. And as far as timeline goes, the Navi dies that are in shipping products this July and the dies that are going into next gen consoles will probably be separated in manufacture dates by an entire calendar year.

 

[Playboi Carti]

What leaks?

Click to shrink...

This one

 

[MrKlaw]

At least I could keep up with the RAM talk. Now yall getting into stuff I can't handle. 😭

Click to shrink...

Wave race 64 launch game for PS5

 

[dgrdsv]

By this time next year AMD will have a 5700 replacement that will get basically the same perf at 120W or less TDP. Mark my words.

Click to shrink...

How will they achieve that?

 

[Miniature Kaiju]

How will they achieve that?

Click to shrink...

Mix of node maturity allowing lower voltages and design optimizations. They've done it before.

 

[disco_potato]

How will they achieve that?

Click to shrink...

Easy. Buy a ton of 2070s and put AMD stickers on them.

 

[Deleted member 1589]

How will they achieve that?

Click to shrink...

 

[Lashley]

Considering that the new generation is starting at about twice the GPU power of mid-gen upgrades of the previous generation, I really doubt that we'll see many 60 fps titles.
People keep thinking for some reason that faster CPU is the only thing you need to get to 60.

Click to shrink...

Yup. Setting themselves up for disappointment.

 

[anexanhume]

Considering that the new generation is starting at about twice the GPU power of mid-gen upgrades of the previous generation, I really doubt that we'll see many 60 fps titles.
People keep thinking for some reason that faster CPU is the only thing you need to get to 60.

Click to shrink...

4K60 is really fucking hard. Very few desktop GPUs can manage that, and it's title dependent then.

 

[Rylen]

How will they achieve that?

Click to shrink...

They've sorta done it with GCN before. And in 10 months time.