Next Generation Hardware Speculation with a Technical Spin [pre E3 2019]

[function]

I've seen a couple of breakdowns on memory use for games, and you're looking at 2~3 GB of the 5 available for games just for textures. And of those 2~3 the majority is going to be texture data you don't need that frame, or probably even for the next few seconds.

With a robust streaming system and a fast SSD, 16GB would probably let you quadruple asset quality with with memory left over for improved render buffers, AI patterns, world simulation etc.

Edit: It's bandwidth rather than huge increases in quantity that are the big concern next gen, I reckon. PS4 Pro is hamstrung by it's bandwidth. X1X gets round it by having a huge bus with a crazy amount of channels, but that's not cheap. If you're more than doubling the Pro's performance (both GPU and CPU) you're probably going to need something beyond 14 Gbpp GDDR6 on a 256 bit bus, and that'll get pricey.

 

[Shifty Geezer]

Since Navi is still GCN, I really doubt it can do significantly more than 33%, unless RPM would be heavily used. Each iteration of GCN was hyped as hell as bringing huge uplifts and all of them were wimps, with uplifts coming mostly from process and clock speeds.

Features like mesh shaders and variable rate shading. We have no idea what the features of Navi (or MS's next console arch) are, so it's silly to compare TFs as the be-all-and-end-all. If a next gen console can look next-gen on a relatively marginal improvement in raw flops, then it's still 'next gen', and we know that's possible. XB1X and the games on it are based on 2013 tech and <2 TF consoles. Next-gen machines will sport 5+ years of conceptual advances and vastly superior baseline targets. The teraflops are a lie. Don't fixate on them.

Historically most consoles were sold at a loss in the beginning.

Historically, most consoles didn't make much money and the companies that operated them went bust...

Exactly and there were discussions already on this forum about that topic and I remember someone (Sebbi?) saying it is very limited. In real world you might use it for a third of the rendering or less, depending on what you are rendering. It's not something you can put a lot of faith on.

Doubling performance for one third of the workload would mean increasing the performance of your GPU by one third for that part of the rendering. In raw flops, that means taking a 9 TF GPU and getting the equivalent of 12 TFs from it. The more devs use it, the more they'll find ways to maximise it, so it really shouldn't be discounted as you are doing. It's one of several features that overall means new GPUs will be able to accomplish more on screen per teraflop of shader power.

 

[function]

I'm pretty excited about seeing Navi Cards drop in Q3 this year. Should give us an idea of how much AMD have advanced in terms of bandwidth bottlenecks. You just know some Uber nerds are going to do some flop/cu/Hz/BW juggling to try and draw comparisons with Vega and Polaris!

 

[goonergaz]

Regarding the TF debate, I can see it being a bit of a PR nightmare for MS going from 6 to 8...not so Bad for Sony as it'll still be 'double the power'.

I do wonder if Cerny laid the groundwork for some 'low TF' console though, all that talk of minimal loading and ray tracing makes me think next gen's battlefield will be based more around those.

As for selling at a loss, as we know the h/w is just one factor so whilst they need to make money the box itself is just an access to ways of getting more revenue.

Edit - sorry it wasn't very technical

 

[see colon]

I'm just doubtful that Microsoft is going to release underpowered hardware next generation. I think the negative pre-launch press IRT physical games and DRM plus the lackluster hardware in Xbox One was a big hit to Microsoft, and I think they see the positive reception the One X has received. I also think they counted on ESRAM and the command processor being the secret sauce that would smooth out performance, but Sony just ate their lunch and drank their juice box. I'm still really underwhelmed by the command processor. I really thought that would give them a CPU advantage that would show in some games, but it never did. There aren't any games that I know of that, when CPU limited, show more than the theoretical 10% advantage that the clock speeds would predict. Unless I missed something.

 

[JoeJ]

I've seen a couple of breakdowns on memory use for games, and you're looking at 2~3 GB of the 5 available for games just for textures. And of those 2~3 the majority is going to be texture data you don't need that frame, or probably even for the next few seconds.

With a robust streaming system and a fast SSD, 16GB would probably let you quadruple asset quality with with memory left over for improved render buffers, AI patterns, world simulation etc.

How will those numbers change with 4K ready textures? And do we need to increase texture resolutions at all for that? There could be surprising HW support for bicubic texture filter to make magnified texels look good enough.

But in any case it would be helpful to have extended compression libraries. Something like jpeg from HD to GPU formats, and you could run it on CPU or GPU like you want, also usable to compress procedural stuff generated at runtime. (Don't know if consoles already have this, but it seems a case where many people work on the same thing.)

 

[Jay]

Doubling performance for one third of the workload would mean increasing the performance of your GPU by one third for that part of the rendering. In raw flops, that means taking a 9 TF GPU and getting the equivalent of 12 TFs from it. The more devs use it, the more they'll find ways to maximise it, so it really shouldn't be discounted as you are doing. It's one of several features that overall means new GPUs will be able to accomplish more on screen per teraflop of shader power.

It was something like 5% improvement in a specific part of frame rendering e.g. 10% of total frame generation. Making up figures but for that use case it wasn't huge.
The benefit was register pressure relief that FP16 gave, RPM is FP16 arithmetic.

But VRS, RPM, mesh shading all added together should add up. But it does entail developers to use it, which I believe isn't a problem, but I know not everyone shares that view here.

Not responding to anyone in particular:
TF is used as a basis of relative performance comparisons, as we have no other way to do it at the moment.
We have no idea of navi, customisations, RTRT implementations etc.
All we can say is if all next gen is navi, base performance is x.

 

[iroboto]

TF is used as a basis of relative performance comparisons, as we have no other way to do it at the moment.

well... it's a generalization with the assumption that the rest of the card is beefed up to support the full saturation of those TF. But that isn't the case often, and many would argue this is definitely often not the case with AMD GPUs thus we often see the argument of AMD flops != nvidia FLOPs.

We have no way to really know what the performance numbers will do or need to be. What we're missing is knowing what MS and Sony are trying to target for their consoles. I don't think the spec numbers matter as much if they are hitting their graphical targets.

that being said, my final guess was Anaconda coming in at 9TF with lockhart at ~6. So after seeing Gubbi's post, i'm going to double down on that.

 

[AlNom]

I've seen a couple of breakdowns on memory use for games, and you're looking at 2~3 GB of the 5 available for games just for textures. And of those 2~3 the majority is going to be texture data you don't need that frame, or probably even for the next few seconds.

With a robust streaming system and a fast SSD, 16GB would probably let you quadruple asset quality with with memory left over for improved render buffers, AI patterns, world simulation etc.

We'll also have a number of architectural improvements for tiled resources and the sort.

Probably should compile a list of items that have been added to D3D FL12_0/current gen.

 

[Jay]

well... it's a generalization with the assumption that the rest of the card is beefed up to support the full saturation of those TF. But that isn't the case often, and many would argue this is definitely often not the case with AMD GPUs thus we often see the argument of AMD flops != nvidia FLOPs.

We have no way to really know what the performance numbers will do or need to be. What we're missing is knowing what MS and Sony are trying to target for their consoles. I don't think the spec numbers matter as much if they are hitting their graphical targets.

I was talking specifically about comparisons in these threads.
It is a generalisation, but it gives the reader an idea of what people think and mean.
Can change TF to potatoes, but TF helps in what people think in comparison to this gen as a general measure.

Only TF I've really settled on is for Lockhart which is 5.5-6TF around that mark, but 6 being absolute max.

 

[Picao84]

I still think that TFlops is a useful metric. Why? Because we are talking about consoles which will have a life span of 7 to 10 years, not graphics cards that can be replaced each 2 years or so. After all, why did Sony and MS launch, for the first time ever, mid generation upgrades with 2x or 4x the TFlops power of the vanilla version?

Yes, you can have lots of good techniques to improve the work the GPU can do with given TFlops. But 3 years down the line, will those be enough to mask a deficit of brute performance when game developers want to push the envelope?

Are we going to see MS launch a low end model, together with a higher end model that modestly improves upon the Xbox One X, for two years later launch yet another more powerful model? A model that will be harmstrung by the lowest common denominator of the low end model? Or will the low end model stop being supported, fragmenting the user base?

I honestly think that the only approach that makes sense now is go big or go home.

 

[iroboto]

We'll also have a number of architectural improvements for tiled resources and the sort.

Probably should compile a list of items that have been added to D3D FL12_0/current gen.

since current gen XBO hardware feature set?
oh man that list is massive.
uh okay, this list may not be 100% accurate but I don't think OG XBO has these, and neither should Scorpio at least supported in FL1_2 Format. But I think if you were just checking which hardware features, it might be easier to just check for D3D12_FEATURE_D3D12_OPTIONS6 <- as this is the latest known one that contains everything listed below. Some of these were rebuilt because of alignment issues across multiple vendors. So xbox may not necessarily have the same features since it's own locked garden.

• DirectX Raytracing (DXR)
• Variable Rate Shading
• View Instancing
• Conservative Rasterization
• Rasterizer Ordered Views
• Programmable Sample Positions
• Depth Bounds Test
• Background Processing
• DirectML

Then from a ShaderModel level.
I think OG Xbox is mainly SM6.
But it may not contain the hardware support for some of those items.
We are now at SM6.4 - which I guess is the model that supports DXR and DML and whatever other little things that needed to be included.
To support SM6.4 there is only 1 bit - so you either support it all or you do not, unlike 6 in which you can support some and not support some.

 

[Xbat]

You seem to forget that MS and Sony needs to make money off of these consoles.

Historically, most consoles didn't make much money and the companies that operated them went bust...

Hmm I wonder why they went with a loss leader approach? It's obvious and It's even more relevant today than it was then. Sony had an amazing year and it's got nothing to do with making a profit on there hardware and everything to do with there eco system.

8 TFlops isn't going to happen. What kind of cu count are you predicting for 8 TFlops.

I believe we going to get 40-44 CU's at 1.8 GHz so between 9-10 TFlops and that's me low balling. I feel 46-48 CU's is possible especially for Anaconda.

This is assuming a $499 price tag and I expect them to take at least a $50 dollar loss.

 

[Shifty Geezer]

I still think that TFlops is a useful metric.

It has limited value. In some cases, like comparing platforms of the same generation launched within a year of each other, they're comparable. For comparing with last gen machines, it's not useful. Hence no need in worrying about how x teraflops this gen versus y teraflops of last gen. Don't give it a moment's consideration and just focus on what's technically and economical possible next-gen when making our console predictions.

 

[shiznit]

Assuming separate CPU die, I'd say no more than 250-275mm budget for GPU/IO. That doesn't leave a ton of room for CUs even without double precision. If they can hit 10TF the marketing people should be happy.

 

[Picao84]

Assuming separate CPU die, I'd say no more than 250-275mm budget for GPU/IO. That doesn't leave a ton of room for CUs even without double precision. If they can hit 10TF the marketing people should be happy.

Vega VII is not that much larger than that at 311mm and it includes 60 CUs for 13.44 TFlops. It would easily fit 48 CUs giving 10.7 TFlops all things equal, so I can't see 10TFlops being out of reach. The only question being power of course and how much Sony and MS budgeted for that. But 10TF is what my expectations are.

 

[shiznit]

331mm, without GDDR interface.

 

[Nisaaru]

331mm, without GDDR interface.

With chiplets that's on the 14nm I/O chiplet anyway and not part of your 7nm GPU, most likely.

 

[Picao84]

331mm, without GDDR interface.

I don't know how much GDDR would represent, but I was being conservative. My calculations could fit up to 52 CUs.

 

[shiznit]

With chiplets that's on the 14nm I/O chiplet anyway and not part of your 7nm GPU, most likely.

Lol. The GPU is the IO die. Are you expecting a ~500GB/s interface between separate IO and GPU dies? Come on.