nial
Member
They have the SoC manufacturing deal in place and that's not going to change. Sony will not sit on these chips for a long time, which means that a delay beyond 3 to 6 months is not feasible.
-
Hey Guest. Check out your NeoGAF Wrapped 2025 results here!
KeplerL2: PS6 Likely to Use 1TB SSD with No Disc Drive; Game Size Could Be Reduced with Neural Texture Compression
- Thread starter IbizaPocholo
- Start date
- Opinion Rumor Hardware
Rocco Schiavone
Member
I have two spare NVME Drives. 4TB and a 2tb. I don't give a shit.
It would be much more important to know whether the games will be released on physical media.
If there are no physical media options, then you might as well just play them on PC.
It would be much more important to know whether the games will be released on physical media.
If there are no physical media options, then you might as well just play them on PC.
Bojji
Member
If game is maxing PS5 SSD, that means it won't run properly on slower storage (pop in, stutters etc.).
So far (6 years into the gen) there are no games like that. Even Ratchet runs fine on Gen3 speeds (on PS5).
Everything gets worse so it might be true!
viveks86
Gold Member
I was referring to R&D and looks like we agree on that.
Tens of millions more devices based on what logic? It's just arbitrary crystal ball gazing. Nobody knows if or when any of these industry-wide issues will blow over. And for all we know it could all settle well before launch and none of this matters. Or not. We have no idea.
See
nial
's post above. This stuff isn't that easy. Component prices are locked in and those deals are not going to change. Volumes may change, but the deals are still contingent upon volume commitments and timing. They are not going to jeopardize all those hard fought negotiations because poor ol' consumer needs to shell out a few hundred more. They'll suck it up and deal with whatever blowback comes from the pricing. Will that affect sales? Definitely. But if the games are there, the generational jump is visible and they market the heck out of it, sales will start slow, but with a much longer tail. And if they are able to bring the costs down over time, then they will get to their original sales targets eventually anyway. This isn't a sprint where launch window is when they sell the most. It's a marathonWait... are you suggesting nobody will release any new product next year? No RDNA 5? No 60 series? No Televisions? For how long are they all pausing so that they don't take too much of our money? And if it's 2 years? Why 2? Why not 5? Or 10?
Last edited:
Buggy Loop
Member
Nvidia's GitHub demo of the technology
viveks86
Gold Member
Haha. No. Not even remotely. These aren't LLMs. These are tiny neural networks called MLPs that you train per texture to replicate the original artwork to the tiniest details (with some loss due to the nature of the technique).
In layman's terms, I show the neural network a super high resolution picture of a leaf, a super low resolution version of the same leaf (called a latent texture) and train it thousands of times so that anytime I give it the latent texture, it returns something really close to the original high resolution.
Once that's in place, now I ship the game with just the MLP and the latent texture, which is around 10% the size of the original texture. Profit!
Alternatively, given the massive reduction in size, you can afford to make really complex, multi-layered textures with effectively the same storage requirements.
The only catch is the runtime performance and unexpected visual issues that arise out of something so new to the graphics pipeline. Both will likely be fully solved next gen. Only question is timing.
Mr.Phoenix
Member
That would be the dumbest thing they could do.
You can't plan the next 7 years of your platform on the off chance that things "might" get cheaper the next year or the year after that. That's just not how these things work at all. And what of if they don't get cheaper, what of if they get even more expensive? So what? You just never release a new console until things get better?
But more importantly.... do you think these OEMs, especially those like sony that are considered Tier 2, go to TSMC, samsung or micron a few weeks before they start mass production of their new console and buy a couple million chips of RAM or storage? Nope. If they plan on launching the PS6 in in november 2027, then volume production would start something in august 2027, they would have contracted the reservation and its price for the components they need by this time. Those contracts are usually made 12-18 months in advance. Those renewed contracts likely has a lot to do with the current price increases for the PS5.
I don't see them not using PCIe 5 SSDs. Its one of those things that are just better to have than not have, especially for something thats going to be on the market for the next 7-12 years. Or for something that 2 years into the life of your console, everything would be PCIe gen 5.
The second the new console is released, and a disc drive is an optional peripheral, Physical media is dead. I don't think next-gen would even give the option of a bundled disc drive SKU. And if thats the case, we would find that even most the physical media purists would put off buying a disc drive add-on alongside the new console..
I even forsee a situation where the physical games will cost more than their digital counterparts, and in some cases even be released later.
viveks86
Gold Member
So this?
A couple of caveats there. I don't believe this version has fully solved the issues with stochastic filtering. You can tell from the intense noise on the green drapes on the right compared to the uncompressed version. And the 0.5ms difference is on a very basic, static scene test on a 5090. Fully animated scenes in-game with complex models and PBR would easily take up a lot more.
The thing about the future though is all this can potentially be done in parallel to other rendering tasks, so the overhead may become a moot point entirely once those issues are solved and neural rendering runs in parallel in all scenarios without fighting for registers and VRAM. But that might be years before mainstream use.
MCH2024
Member
Engineering and System design is about working with constraints.
J Required
Banned
Any spec is limited when utilized fully by a capable developer. This is how the market moves on.
kazuyamishima
Member
Yes, it was, it took a bit more than 6 months I reckon to have decent stock.
twilo99
Member
Should've happened a long time ago
Buggy Loop
Member
Not that one I had in mind, it was an older demo but that works too, even shows that latency wise for a bigger scene than just the helmet
They have, it's even a requirement of NTC and one of the "drawbacks" of it as it does not use hardware filtering, it absolutely needs STF. NTC gives one pixel per decode. Trilinear or aniso would be too expensive.
GitHub - NVIDIA-RTX/RTXNTC: NVIDIA Neural Texture Compression SDK
NVIDIA Neural Texture Compression SDK. Contribute to NVIDIA-RTX/RTXNTC development by creating an account on GitHub.
github.com
there's a bug with the reference actually. All modes in the demo have STF.
We don't know that in the end. We have no idea of the occupancy of the GPU for this. There's a fixed cost sometimes for inference that it won't budge no matter how much NPUs you throw at it. DLSS comes to mind, its impacted of course to some degree but nowhere near what a 2060 to 5090 TOPs difference would show. When you get into the ~0.5-1ms frametimes for an effect in a pipeline inference, you might be capped there for a long time no matter how many tensor cores you throw at it.
The thing about the future though is all this can potentially be done in parallel to other rendering tasks, so the overhead may become a moot point entirely once those issues are solved and neural rendering runs in parallel in all scenarios without fighting for registers and VRAM. But that might be years before mainstream use.
Tensor cores are always working concurrently of course, on top that these can be decoded independently and even in the Gbuffer field or ray tracing passes.
They also talk about it in their papers
That's why I said the performance impact of the small initial demos are not really representative of a full game frametime. There's a shitload of things in a modern game frametime that you can insert this concurrently with minimal impact.
There's also a mode in NTC that will use sampler feedback streaming, in a future DX12 implementation.
Inference on feedback is a middle ground between on-sample and on-load. It reduces VRAM a lot, not as much as on-load, but also better performance.
And as Kepler says, inference on load is basically a free lunch, you get the benefits of disk space and bandwidth, VRAM usage is still a problem though. Which is why I think everyone will look at inference on sampling with the prices of VRAM nowadays
There's an explosion of papers on neural compression and not just textures, there's even papers on how to apply neural compression to grids of light probes to have global illumination maps baked for different times of day or states of environment.
Oh also just Sampler feedback streaming, when it was implemented in half life 2 RTX it has huge VRAM savings, even before NTC.
Last edited:
viveks86
Gold Member
Not that one I had in mind, it was an older demo but that works too, even shows that latency wise for a bigger scene than just the helmet
They have, it's even a requirement of NTC and one of the "drawbacks" of it as it does not use hardware filtering, it absolutely needs STF. NTC gives one pixel per decode. Trilinear or aniso would be too expensive.
GitHub - NVIDIA-RTX/RTXNTC: NVIDIA Neural Texture Compression SDK
NVIDIA Neural Texture Compression SDK. Contribute to NVIDIA-RTX/RTXNTC development by creating an account on GitHub.
there's a bug with the reference actually. All modes in the demo have STF.
We don't know that in the end. We have no idea of the occupancy of the GPU for this. There's a fixed cost sometimes for inference that it won't budge no matter how much NPUs you throw at it. DLSS comes to mind, its impacted of course to some degree but nowhere near what a 2060 to 5090 TOPs difference would show. When you get into the ~0.5-1ms frametimes for an effect in a pipeline inference, you might be capped there for a long time no matter how many tensor cores you throw at it.
Tensor cores are always working concurrently of course, on top that these can be decoded independently and even in the Gbuffer field or ray tracing passes.
They also talk about it in their papers
That's why I said the performance impact of the small initial demos are not really representative of a full game frametime. There's a shitload of things in a modern game frametime that you can insert this concurrently with minimal impact.
There's also a mode in NTC that will use sampler feedback streaming, in a future DX12 implementation.
Inference on feedback is a middle ground between on-sample and on-load. It reduces VRAM a lot, not as much as on-load, but also better performance.
And as Kepler says, inference on load is basically a free lunch, you get the benefits of disk space and bandwidth, VRAM usage is still a problem though. Which is why I think everyone will look at inference on sampling with the prices of VRAM nowadays
There's an explosion of papers on neural compression and not just textures, there's even papers on how to apply neural compression to grids of light probes to have global illumination maps baked for different times of day or states of environment.
Oh also just Sampler feedback streaming, when it was implemented in half life 2 RTX it has huge VRAM savings, even before NTC.
Yes the demo has STF, but STF itself has lots of room for improvement and not in its end state.
And there is no need for STF if the inference can produce a pre filtered output. See Ubisoft's paper on that. You could get all the benefits of NTC without need for filtering separately.
https://arxiv.org/pdf/2311.16121
And inference on load saves disk space, but not VRAM bandwidth. Once the texture is loaded in memory as BC, it's no different than regular BC textures, so all rendering calls will use the same VRAM bandwidth as without NTC.
But I do agree that inference on sampling is the right approach forward. I feel that until that gets mainstream adoption amongst devs, the half step "free lunch" may not catch on as there are lots of workflow changes that need to happen to allow for quick edits, retraining etc. and on top of that it only resolves the smaller bottleneck (SSD space) without solving the bigger bottleneck (VRAM bandwidth).
Agree with the rest. Using neural rendering for baking GI seems like a cool idea that I had not read about. Will study that.
Last edited:
Mr.Phoenix
Member
I honestly can see why sony chose to go with Universal compression, if doing NTC, you can't do inference on sampling; there is fundamentally no need to do it outside of having a smaller game size and only needing to move smaller amounts of data into RAM. But none of those things actually helps with actual RAM and bandwidth utilization.
The good thing, though, is that ist ultimately just an AI operation, and the PS6 will have the hardware to be able to do it if its ever developed enough to be adopted.
Actually, not nuts at all... what would be nuts is if they didn't give you the option or ability to upgrade your drive yourself. The truth here is that having an SSD at all is a luxury, lets be real here. Once upon a time, games came on carts or discs, you could only play one game at a time, and if you wanted to play a difffernt game, you got up and switched it... and don't get me started on the loading times on some of those dics.
All they have to do with an SSD is realistically give you enough space to have like 4-5 big games installed at any given time, you are free to delete and install other games as you see fit, if you want more space, buy an expansion drive. Hell, you can even use an old-fashioned HDD as a cold storage drive if your issue is having to redownload games.
Another way of looking at this is, if they asked you, to choose 4TF more or 5GB more of VRAM for the life of your console and can never upgrade it, or choose 1TB extra of storage space that you can always upgrade yourself anyways... which would you go with?
What you are basically saying here is, why does the system only allow me to install 8-15 games at a time? I want to be able to install double that amount out of the box.
viveks86
Gold Member
I think inference on sampling would absolutely be achieved. It would almost be a trivial task with next gen hardware across the board, in both the PC and console space. It's just a matter of timing and readiness for launch.
Universal compression is taking a more generic approach that could apply on all data types (and not just textures) with focus being vram bandwidth and not storage.
But a generic approach naturally comes with shortcomings. It won't remotely compete with the compression ratios that NTC can unlock. Unless it's a "Swiss Army knife" suite of compression tech, that picks the appropriate method for the given data type. In that case, it could leverage NTC for textures while using meshlet compression for geometry, or compress DGF even further? I don't know. Hard to say anything about UC without any literature on it. I don't think there is even any patent published on it so far (or i might have missed it). So it's difficult to guess what they are doing.
But given one is focused specifically on VRAM bandwidth and the other is focused specifically on textures, I think the chances that both would work in tandem are high. That gives you wins on all fronts. Like K KeplerL2 said, they aren't mutually exclusive.
viveks86
Gold Member
Hopefully, yes. In your own quote it says
All this needs to be set in stone with clear findings before devs take these and run with it. Because once you leave this station, there is no coming back. AC Mirage is already dipping its feet to some degree, but they had to hold back on a complete overhaul due to performance costs. What they did is awesome from a research standpoint, but cost of production is technically going up by hand picking what to compress how.
https://www.ubisoft.com/en-us/news/...exture-compression-in-assassin-s-creed-mirage
I'm sure this will all be figured out in due course. Just hard to rush to a conclusion on production readiness based on early testing.
Nvidia, and intel to some degree, are well ahead of the game on these aspects with dedicated tensor cores and cooperative vectors. AMD is playing catch up as always. And in context of the PS6, that can't be ignored. Hopefully it all works out. Everything we have heard so far is well inline, if not ahead, of Blackwell. So that's good news!
Last edited:
RPS37
Gold Member
Literally fits 10 games.
Can they please just give us more space to start with already?
Last time I checked it was $550 now.
Can they please just give us more space to start with already?
Last time I checked it was $550 now.
coffinbirth
Member
I would.
Would absolutely NUKE their retail presence, and give retailers ZERO reason to stock the consoles.
They would have to shift profit margins from software to hardware, which would already be sold at break-even, or potentially at a loss, and given the current situation they have with component pricing, they would be shooting their dick completely off.
Still very much a "have the razors to sell the blades" scenario.
Plus, they are still going to be selling PS5 games for the next ten years...and those disc drives are eeeaaasy money.
Not to mention, the backlash would be immense.
Guardian Monkey
Member
This is silly. Why bring the console out if they are going to chop it apart before launch. Smaller hard drive, less ram, the system is not a generational leap from the pro even. Especially with it mainly leaning on pssr and frame gen this is going to age quick. Just wait till there is hardware that warrants it if they are resetting the gen instead of rushing it out underpowered and overpriced. Take the loss on the dev cost and start again if need be. The time isn't right if they rushing it out to just get it out.
MCH2024
Member
SSD speed isn't one of them
HerjansEagleFeeder
Member
But they have a financial incentive not to sell enough disc drives as their desperate early adopters will just buy digital with higher profit margins then.
Panajev2001a
GAF's Pleasant Genius
Do not tell me we are back to the early PS2 days with sloped surfaces and blurryness
.Oh well, we will see. We need to do something to combat storage and memory space issues, just wish we did not have to hehe.
Mr.Phoenix
Member
I believe UC is just a next-gen version of what they did with kraken on the PS5. IV basically allows the PS6 to make 30 GB of RAM appear and function as 45GB of RAM. Numbers from my ass but in theory that's how its supposed to work. And yes, I agree that both can work together.
The fact you say stuff like the system's not a generational leap over the Pro...and then basically saying that it having neural arrays will mean it will age quick... actually speaks volumes about how uninformed you are about stuff like this and where tech design is going in. You are likely one of those people who still think TFs is how you measure generational leaps. While ignoring all the key technologies that they have been talking about and have leaked that should actually get you excited about the next gen... if only you understood it.
Then again, I expect nothing less from a clear example of the Dunning-Kruger Effect. Underpowered and overpriced... lol.... You couldn't be more wrong if you tried.
At some point in every tech industry... more just wouldn't cut it anymore... You have to start building smarter.
Last edited:
Perrott
Member
Forget about cartridges: the handheld won't support any form of physical media.
That's the play: having you fully captive to the PlayStation Store. The handheld form factor is just the Troyan horse.
Guardian Monkey
Member
Cumulus
Member
lol, talk about being completely delusional. The only GPU that has a chance of dethroning the 5090 in the foreseeable future will be the 6090 and then, the 7090 after that (if we`re lucky the 7080, too). The 5090 will most likely be in the top 3-4 fastest GPUs until ~2030 at least.....and its price will reflect that.
Last edited:
HeisenbergFX4
Gold Member
Mr.Phoenix
Member
Unlike some people, I won't take research and data and make it sound like its all mine.... You should consider doing some research, though. The material is there for you to see for yourself. Sure as hell would beat you saying fuck all and making general assumptions based on nothing but your feelings."Numbers from my ass"
FauxXeno
Member
I haven't bought physical media since 2015 so it makes no difference to me , I'll still get access to everything I've bought for ps4/ps5 . If anything the physical media andy's are held captives into spending more upfront for new consoles and there will be a time when Sony finally throws in the towel on it and you won't be able to take advantage of future hardware and/or patches/updates and locking you out of your physical library completely . By the time ps7/8 comes out the only way to play your collection will be dusting off the old ps4/5/6 lol . Games are getting so big we're re-approaching the ps1 era need of having multiple BR discs or a mandatory downloads to play the game while Sony fails to futureproof the medium . While I have no stake in physical media I just think it's time to evolve especially when in modern times a small cartridge can hold x10 the data and will only act as a authentication key and install
MCH2024
Member
Compression is not new. And GPUs already do it.
Lets shut this kool aid narrative now. Kracken does 5.5 GB/s => Up to 9 GB/s. Aka that's how much it can decompress, 5.5 GBs every second with data that's compressed to 61% of original size at most.
you're asking for Kracken to now do the same thing, except now it has to decompress 640 GB/s (120x) …….
The CPU is roughly a 9600X with worse single core. But you basically sacrifice an entire core to do 6-7 GB/s of very fast decompression. (Way worse than Kracken's compression, by the way)
And even if it was possible to build such an ASIC. You can't add latency to memory without degrading GPU performance. Not on a GPU with so little cache as PS6.
This is just hocus locus stuff. Sorry.
Last edited:
CentralScrutinizer
Member
I am going to use my imagination and pretend that 'Neural Texture Compression" is the game shipping with missing textures and just let AI make it up along the way DLSS 5 style.
Bojji
Member
So when I played this the first time, I noticed that Maria's eye colors were different when the two Pyramid Heads knock her off near the end of the game. I'm making a YouTube video on this game, so I wanted to include this. I went to go take a screenshot from the video and much to my horror...this was not a miscolored eye...
IT'S FUCKING TEETH AND GUMS IN HER RIGHT EYE.
WHAT. THE. FUCK. :LOL
IT'S FUCKING TEETH AND GUMS IN HER RIGHT EYE.
WHAT. THE. FUCK. :LOL
EverydayBeast
ChatShitGPT Alpha 0.001
It's time to end discs and most people will say "finally".
Bojji
Member
I guess it's time to get fucked in the ass by Sony.
DE only console with one selling vendor is the worst thing possible.
Guardian Monkey
Member
Teraflops do matter, ram does matter. Computationally it's not a generational leap above the pro, I still stand by that, and so do a lot of people pushing back at the idea of it coming out next year. You can't just AI this shit away saying the fundamentals don't matter. Especially when we haven't seen rdna5 in full effect and this hardware will be the first gen of it.
Topher
Identifies as young
adamsapple
Or is it just one of Phil's balls in my throat?
So when I played this the first time, I noticed that Maria's eye colors were different when the two Pyramid Heads knock her off near the end of the game. I'm making a YouTube video on this game, so I wanted to include this. I went to go take a screenshot from the video and much to my horror...this was not a miscolored eye...
IT'S FUCKING TEETH AND GUMS IN HER RIGHT EYE.
WHAT. THE. FUCK. :LOL
I fucking hate what Konami did to SH2's legacy with this "remaster".
HeisenbergFX4
Gold Member
Why would most people say "finally" to options being taken away?
MCH2024
Member
It's a half gen leap. PS5 Pro is Gen 9.5 (Turing, Ampere, RDNA4). PS6 is Gen 10 (Ada, Blackwell, RDNA5).
The biggest difference is path tracing. Ray tracing and rasterization are like 75% uplifts.
Last edited:
DavJay
Member
Not crazy that they only give us a TB but crazy to the fact how slow storage space are growing. The 360 went for 20GB and by the end we were at 500GB. You'd think we would be at 50TB as a standard by now.
Ai plays a part but still damn. So grateful for the 2TB PS5Pro. That was more of a selling point to be than the raw power.
Audiophile
Member
Sony could always go for a middleground with a 6Ch or 12Ch setup (the latter being what they did with the launch PS5) and target 1.50TB or 1.65TB. Assuming complexity doesn't outweigh cost reductions on capacity.
If they user 'NTC' + improved general compression they could in the real world cut data down to 2/3 while games double quality and double quantity/variety of assets at the same time. With native titles his would be equivalent to a little more than 2TB storage in current console configurations on the aforementioned 1.5-1.65TB config.
As for PCIe gen, I doubt there's much reason to limit it to Gen 4. I don't think the cost savings over the gen for 4 vs 5 will be all that significant. Better to go for a lower end PCIe5 in the 9-11GB/s RAW range imo. I still think there's room for paradigm shifts on that front and it'd be a mistake to give up on it just because devs didn't jump on straight away, this last gen was bogged down by prolonged cross gen with those old mechanical HDDs having to be considered and those practices even being carried over to the first run of current gen only titles. But moving forward with 2.5-5.5GB/s RAW SSDs being the baseline for the upcoming cross-gen period, they'll weigh things down far less while a ~2-4x effective base speed over that could provide a nice relative bump if some devs try to change up how things are done now that engines are fully evolving away from those HDDs.
If they user 'NTC' + improved general compression they could in the real world cut data down to 2/3 while games double quality and double quantity/variety of assets at the same time. With native titles his would be equivalent to a little more than 2TB storage in current console configurations on the aforementioned 1.5-1.65TB config.
As for PCIe gen, I doubt there's much reason to limit it to Gen 4. I don't think the cost savings over the gen for 4 vs 5 will be all that significant. Better to go for a lower end PCIe5 in the 9-11GB/s RAW range imo. I still think there's room for paradigm shifts on that front and it'd be a mistake to give up on it just because devs didn't jump on straight away, this last gen was bogged down by prolonged cross gen with those old mechanical HDDs having to be considered and those practices even being carried over to the first run of current gen only titles. But moving forward with 2.5-5.5GB/s RAW SSDs being the baseline for the upcoming cross-gen period, they'll weigh things down far less while a ~2-4x effective base speed over that could provide a nice relative bump if some devs try to change up how things are done now that engines are fully evolving away from those HDDs.
Last edited:
viveks86
Gold Member
Wait, why are you making this seem like it's unachievable? Blackwell already has a Decompression Engine ASIC that does 800 GB/s throughout.
This sounds like an extension of that idea, but for graphics workloads (while DE seems tuned for AI). This would remove the need for CPU side decompression altogether. So not sure what you mean by dedicating 1 CPU for it.
This has nothing to do with Kraken and the IO complex. I think that was used just as a metaphor. This is within the GPU and there is already precedence for it from Nvidia.
Do you have an alternative explanation on what UC is claiming to be? Or do you think it is just vaporware? It's clearly not just Delta Color Compression. So what even is it? Let's define it before discounting it.
Similar threads
- Poll