Next Generation Hardware Speculation with a Technical Spin [post E3 2019, pre GDC 2020] [XBSX, PS5]

Status
Not open for further replies.
PS4 had only about 160 GB/s of maximum effective bandwidth. 530 GB/s is 3.3x more. I wouldn't worry about it. It's going to be enough.
The difference between physical hardware memory b/w and the real world applies to any SoC. So unless AMD did some crazy magic, that 530 would be proportionally lower too. Apples to apples is 530/176. It's still 3x so probably okay because the OG ps4 didn't have compression, the ps4 pro is a better comparison. Also RDNA probably have improvements too?

We're seeing about 50GB/s of memory bandwidth per TF in current pc gpus and mid-gen consoles alike. It looks like the balance they would aim for.

I don't think we will get 18gbps as a mid-grade bin next year. Is that 530GB/s confirmed? Maybe 320bit or something.
 
Last edited:
I have been consistently surprised by what the platform holders are delivering for this gen. I'm out of the speculation game. I'll stick to reactions on what gets revealed. If anything illustrates how little we, as outsiders, can possibly know about the process of delivering these console designs, it's how poorly we were able to predict what was actually coming.
 
You cannot replace a big Ram Amount with streaming via SSD.
SSDs help with streaming, they do not help with data that must be kept available. Do you want ray tracing? You need ray tracing data structures in RAM. Want better lighting? Requires RAM for GI probes or voxels. Better shadows? Requires RAM for larger shadow cache. Better BRFS? More RAM for fatter G-buffers. Better AI? More RAM for game code. All of this cannot be streamed and is independent of the resolution. So yes Next Gen Progress , Game Development is holding back with the 16 Gbyte Ram limitation, and the lowest common Dominator called "Lockart".

Sure. An SSD won't help with simulation data and such stuff. But by freeing up memory from the stuff you CAN stream from the SSD, devs do get a larger partition of ram to be used by the things that can't be streamed.

There is a good way to think about how much RAM is actually used by dynamic stuff or needed to be read every frame, and how much is just cached there for if and when it's needed. Suggested by Sebbbi no less. No matter how much total RAM a system can have, you can only show as much the stuff stored there as the bandwith allows within a frame.

The PS4, for example, while affording games with 5Gbs of RAM, has a bandwith of about 180Gb/s. Divide that by 60fps, and you are left with 3Gb useable per frame TOPS. That's ignoring the fact that real world code will never use all that bandwith optimally. So one could say a real world game, even a well optimised one, will make use of no more than 2 fo 2.5Gb of data per frame. Everything else in there is just cashed for later use.

That means more than half of the data sitting in RAM in any 60fps ps4 game could just as well had been in the Hard Drive and be streamed just in time only when needed to, if only its Hard Drive had been an SSD. Well, that's gonna be the case with ps5 now.
 
Last edited:
Sure. An SSD won't help with simulation data and such stuff. But by freeing up memory from the stuff you CAN stream from the SSD, devs do get a larger partition of ram to be used by the things that can't be streamed.

There is a good way to think about how much RAM is actually used by dynamic stuff or needed to be read every frame, and how much is just cached there for if and when it's needed. Suggested by Sebbbi no less. No matter how much total RAM a system can have, you can only show as much the stuff stored there as the bandwith allows within a frame.

The PS4, for example, while affording games with 5Gbs of RAM, has a bamdwith of about 180Gb/s. Divide that by 60fps, and you are left with 3Gb useable per frame TOPS. That's ignoring the fact that real world code will never use all that bandwith optimally. So one could say a real world game, even a well optimised one, will make use of no more than 2 fo 2.5Gb of data per frame. Everything else in there is just cashed for later use.

That means more than half of the data sitting in RAM in any 60fps ps4 game could just as well had been in the Hard Drive and be streamed just in time only when needed to, if only its Hard Drive had been an SSD. Well, that's gonna be the case with ps5 now.

With virtual texturing you don't need tons of memory dedicated to texture in RAM by our own sebbbi. 256 MB for textures in RAM is great or a bit more depending of decals...

https://forum.beyond3d.com/threads/...ecome-more-imporant-again.58288/#post-1934072

Id-software is using virtual texturing. VT has (almost) constant texture memory cost, no matter how high texture resolution or how many different textures you have. With VT you can basically use all your assets in every single level, and VT only loads those texels that are currently visible. As there are fixed amount of pixels in the screen (~2M for 1080p, ~8M for 4K) there is a hard limit of texels you need to have loaded.




For example our Xbox 360 games (~720p) had a fixed size 48 MB VT texture atlas in memory, and that was used to store all loaded texture data (except for some special cases). Players could create levels with dozens of 4K textures visible simultaneously, but the VT system would never use more than 48 MB of memory. This works fine, as each new visible surface covers as many pixels as it adds, roughly maintaining a constant 1:1 visible texel ratio. Virtual texturing doesn't load texture data for hidden surfaces, so overdraw doesn't count.




With virtual texturing, a single 256 MB texture atlas should be enough (to hold all texture data) for a modern PBR pipeline at 4K. Additional texture data might however be needed depending on how you do your decaling and how you composite your materials at runtime. Id-sofware is nowadays using "dynamic" virtual texturing, meaning that they no longer just load baked data. VT pages are generated at runtime (similar to the RedLynx system). This is a great way to reduce the disk storage, and amortize decaling and material compositing, etc cost over multiple frames.
 
With virtual texturing you don't need tons of memory dedicated to texture in RAM by our own sebbbi. 256 MB for textures in RAM is great or a bit more depending of decals...

https://forum.beyond3d.com/threads/...ecome-more-imporant-again.58288/#post-1934072
Which is milk's point, only he was presenting a far more conservative sort of maths that doesn't assume tiled resources. You don't need many, many GB of RAM if you have fast storage as the assets actually renderable are very limited.

I suppose Karmaprof's very pessimistic POV would say that in the world of ray0tracing, you are no longer limited to sampling pixels and data requirements will be massively more, needing to sample all the geometry of a scene rather than just that within the viewing frustum.
 
Alternatively, it's still almost a year away from launch and Microsoft might still be waiting to see what the final silicon will do as they get closer to when they'll start manufacturing the consoles for launch.

They likely have a ballpark idea of what the SOC will do, but will get an even better idea when they start getting 100's of thousands of chips rather than the limited quantity they are testing with now.

Similar to how the XBO SOC got a late uplift in CPU frequency relatively close to launch.

Regards,
SB
CPU and GPU actually. And yeah its an interesting thought. While rumours suggest 12TF that might just be what they can guarantee from early silicon sample and might very well increase. They might have a bit more leeway from to clock higher. This goes for both manufacturers btw.
 
I suppose Karmaprof's very pessimistic POV would say that in the world of ray0tracing, you are no longer limited to sampling pixels and data requirements will be massively more, needing to sample all the geometry of a scene rather than just that within the viewing frustum.

Yeah, raytracing is the real Memory/bandwith killer. That's why I'm anticipating severe LODing for the RT representation of the scene.
 
Wait. Now PS5 is 24GB of mem... 8GB HBM 16GB DDR4. These claims are getting ridiculous.

Just wait even more, things will even get more ridiculous on the sony side.

You cannot replace a big Ram Amount with streaming via SSD.
SSDs help with streaming, they do not help with data that must be kept available. Do you want ray tracing? You need ray tracing data structures in RAM. Want better lighting? Requires RAM for GI probes or voxels. Better shadows? Requires RAM for larger shadow cache. Better BRFS? More RAM for fatter G-buffers. Better AI? More RAM for game code. All of this cannot be streamed and is independent of the resolution. So yes Next Gen Progress , Game Development is holding back with the 16 Gbyte Ram limitation, and the lowest common Dominator called "Lockart".

I don't believe the 16GB, and just 13GB to games rumor, earlier leaks showed 20GB ram.

If anything illustrates how little we, as outsiders, can possibly know about the process of delivering these console designs, it's how poorly we were able to predict what was actually coming.

So true :)

While rumours suggest 12TF that might just be what they can guarantee from early silicon sample and might very well increase. They might have a bit more leeway from to clock higher.

15TF should be possible late next year.
 
Within the last year, most predictions on b3d have been narrowed down to:
2019 or 2020
Navi 10TF +/- 20%
16GB to 24GB gddr6 with some hoping for hbm (or mixed)
Either 1TB SSD or hybrid
8 or 12 cores zen2 with a range of clock choices

I think we did much better than previous generations. Not perfect, but consider the other gen and mid gen had a 40% difference between both brands, we had a pretty good ballpark of what was possible. It's all about AMD having RDNA ready.
 
CPU and GPU actually. And yeah its an interesting thought. While rumours suggest 12TF that might just be what they can guarantee from early silicon sample and might very well increase. They might have a bit more leeway from to clock higher. This goes for both manufacturers btw.

No. They didn't do it for the 1X they won't do it for Sex. Both 6 and 12 were the highest end of the target ranges. 5-6 for Scorpio and 10-12 for Anaconda.
 
The difference between physical hardware memory b/w and the real world applies to any SoC. So unless AMD did some crazy magic, that 530 would be proportionally lower too. Apples to apples is 530/176. It's still 3x so probably okay because the OG ps4 didn't have compression, the ps4 pro is a better comparison. Also RDNA probably have improvements too?

We're seeing about 50GB/s of memory bandwidth per TF in current pc gpus and mid-gen consoles alike. It looks like the balance they would aim for.

I don't think we will get 18gbps as a mid-grade bin next year. Is that 530GB/s confirmed? Maybe 320bit or something.
No. Apples to apples it's 576 / 176 (or 530 / 160). About 530 GB/s was the alleged tested bandwidth in the test. 160 is the max bandwidth tested on PS4 that I know.
 
No. Apples to apples it's 576 / 176 (or 530 / 160). About 530 GB/s was the alleged tested bandwidth in the test. 160 is the max bandwidth tested on PS4 that I know.
Oh that's right, sorry I thought it was the specs instead of a benchmark.
 
No. They didn't do it for the 1X they won't do it for Sex. Both 6 and 12 were the highest end of the target ranges. 5-6 for Scorpio and 10-12 for Anaconda.
I meant the OG Xbox One. It had a clock speed bump after the original specs were announced. As for the 1X, according to Albert Panello when the chips started coming in they got some with all 44CUs effective, giving a TF count of 6.6TF, which they decided to use as dev kits.
 
Within the last year, most predictions on b3d have been narrowed down to:
2019 or 2020
Navi 10TF +/- 20%
16GB to 24GB gddr6 with some hoping for hbm (or mixed)
Either 1TB SSD or hybrid
8 or 12 cores zen2 with a range of clock choices

I think we did much better than previous generations. Not perfect, but consider the other gen and mid gen had a 40% difference between both brands, we had a pretty good ballpark of what was possible. It's all about AMD having RDNA ready.

Last gen it was handed to us in leaked docs. Other than some tin-foil hat stuff, there wasn't much to speculate on.

This gen RT was very much in doubt (until it was officially announced). Navi being more efficient in applying it's theoretical performance to actual games workloads was hoped for but, given AMD's recent performance wasn't considered a given. And when this was shown, it lead many (myself included) to revise our expectations downward since a configuration with fewer TF could deliver the prior expected performance. Some were skeptical of SSDs being possible due to cost and storage requirements for next gen. Zen2 was expected, but some wondered if it would be stripped down and I don't think many saw 3.5GHz coming. We hit the baseline of what RAM could be expected, but we topped out at 12 last gen and I think 16 was pretty much the least that could be expected (for the performance tier, anyway). Many expected, even demanded, more.

Most of all, though, while some individual specs are more or less of a departure from the consensus, I don't think anyone thought we would be getting all of these things. One spec higher meant one spec had to be compromised. But, ultimately, the only thing that you could argue was compromised was the RAM amount. To me, that is the surprise. Console hardware is powerful again, in absolute terms, and that's pretty exciting. I wasn't sure they had it in them.
 
many saw 3.5GHz coming

3.5ghz confirmed? It is a mere 300mhz more then the 3.2ghz most thought it would be. Kinda low clocked compared to R7 variants, not too strange.

I don't think anyone thought we would be getting all of these things.

Without an nvme SSD, it wouldn't be another gen for me. RT has been here since 2018, a given it would land in 2020 devices, even non-consoles. Zen 3 will be there 2020, atleast zen 2 in there isn't too strange either.
I think the specs are in line of what a modern late 2020 console should have. 13/14TF turing in 2018, 18TF wouldn't be too far off for the high end RTX3000/rdna2.0.

Without the specs the consoles got, they wouldn't be mid range but lower then that.
 
Most of all, though, while some individual specs are more or less of a departure from the consensus, I don't think anyone thought we would be getting all of these things...
The final analysis should be reserved until we know pricing and losses. When entertaining a $500 price, or notably lossy hardware, those possibilities were definitely part of the B3D gestalt. All these things in a $400 box without incurring significant losses would mean B3D grossly overestimated the limits facing next-gen production.
 
The final analysis should be reserved until we know pricing and losses. When entertaining a $500 price, or notably lossy hardware, those possibilities were definitely part of the B3D gestalt. All these things in a $400 box without incurring significant losses would mean B3D grossly overestimated the limits facing next-gen production.

$500 + losses, maybe, but fair point. Many things become possible if you throw enough money at them.
 
Status
Not open for further replies.
Back
Top