Xbox Series S [XBSS] (Lockhart) General Rumors and Speculation *spawn*
- Status
I've never subscribed to that interpretation.
I believe the 100GB was just an example, and is just the game package or part of it, and the way you can choose to access it.
So the engine can choose to access the assets as if it was extention of memory not a separate cornered off 100GB section of the ssd.
I just think this would be pretty hard, as the game would be expecting to be able to access 9GB of memory and swapping in and out at a system level invisible to the game engine sounds like it would be prone to misses, and those misses would have a huge impact on the game. But if possible, I'd be down with that.
1X version of the games would give some distinct advantages, settings, graphics and performance modes etc.
Think a lot of the really bad image quality of the XO is down to resolution though, even makes the same textures look bad. Extremely high AF, AA, double resolution could go a really long way.
Don't disagree, only thing I'd say though is unlike Nvidia having to do it on pc. You know how much TF you have to dedicate to it in this circumstance.On the one hand in a basic BC mode the next gen consoles should have a lot of power left over, but on the other hand if regular compute shaders could do this fast enough, I don't see why Nvidia would have stuffed their GPUs with Tensors. Though over time techniques do tend to become better, faster and less expensive.
So ... basically ... I have no idea.
3TF.
Last edited:
Whether the 100GB is a literal amount (it might be representative of a limit for addressable range per game) I agree it's probably not a separate partition. The thought that struck me was just how good a way this is to make good use of the SSD and to make it as transparent as possible. Paging on demand means you naturally tend to spread out accesses due to the way game worlds exist in 3D space (data is physically spread out around the game world), it means you probably keep latency low automatically, and it means you only transfer what you need or are very likely need (the ideal with any type of transfer system).
It fits very well, from what I understand, with the idea behind sampler feedback and SFS. Infact keeping textures in virtual memory and using SF/SFS to automate the process of fetching and blending seems like a really good fit.
I think running games on Lockhart will turn out remarkably well given the memory differences - especially if the virtual memory is of the same size and works the same way. It strikes me as a very scalable approach!
I wasn't really thinking about X1 BC there (but it's not beyond the realms of possibility IMO), sorry if that wasn't clear!
I was thinking about managing both Lockhart and Anaconda versions of a next gen game and minimising developer considerations wrt to the memory setups.
Hard to say, Nvidia use very powerful Tensor cores to do that, and I have no idea how it would fare on regular compute shaders. Tensor cores seem to support a range of int and float formats, and while the XSX does support accelerated int8 and int4 rates I don't know if these could be used or if they'd be fast even if they could.
On the one hand in a basic BC mode the next gen consoles should have a lot of power left over, but on the other hand if regular compute shaders could do this fast enough, I don't see why Nvidia would have stuffed their GPUs with Tensors. Though over time techniques do tend to become better, faster and less expensive.
So ... basically ... I have no idea.
In my opinion, the hardware-accelerated int8 / int4 method used in XSX can perform highly efficient ray-tracing. This has also been confirmed from developer sources. The same hardware unit is likely to work well for DirectML upgrades in XSX.
Yea, I agree with you I was still stuck in thinking about BC.
10GB sounds like a extremely low amount especially as my ongoing view has been 12GB.
But for next gen games, my view is that Sony and MS know what their doing until proven otherwise.
So if 10GB with SSD was deemed enough I'm inclined to believe it will be. I don't think MS would put themselves in the same situation as they found themselves with the XO.
With the lower texture sizes, buffer sizes, framebuffer size, as long as the GDDR6 bandwidth is high enough and ssd, I don't see it being a problem.
It will take a bit more texture handling by the studio though, the bigger the difference the more work will need to be required.
yea get ya now.
Well MS ML solution is also supposed to be performing LOD management in the Velocity Architecture as well as any upres so whatever ML the Series X has it will be in constant use it would seem.
I’m excited for the potential; but I don’t know the costs or quality. So I’m just waiting to see something before making further statements.
Interesting, thank you!
I wonder if this could mean pre-production validation and testing is now done and they have been able to see that the Lockhart thermals look good without a downclock?
Yup
I am doubt they'll tweak the CPU clocks without looking at other parts of the system too. For Durango, they bumped both the CPU and GPU.
This bit is interesting regarding streaming...
What's Dante & Edinburgh?
At least I saw no more LockharD. LOL
Tommy McClain
Nested ifs.
Yuck.
The cpu in these consoles sip power compared to the GPU portions. Lockhert even if they do manage to get up to 5 or 6 tflops like I was told they are testing for would use close to 1/3rd the power of XSX
RIP portable Lockhart dream. I was hoping that at 3nm they could make a portable Xbox using the Lockhart APU but that isn't happening with a 3.6GHz click on the cpu. 3.2 ghz with a 4tflop gpu would've been perfect. Ah well.
Hmm so it’s still debatable if it will release. This likely is coming down to what PS5 is priced at I imagine.This bit is interesting regarding streaming...
What's Dante & Edinburgh?
At least I saw no more LockharD. LOL
Tommy McClain
3nm is years out. Are you familiar enough with the characteristics of the node and Zen to write a portable off based on a 13% upclock?
I'm aware. It was 2023 at the earliest but COVID has probably pushed that back to 2024 or even 2025.
Are you familiar enough with the characteristics of the node and Zen to write a portable off based on a 13% upclock?
Looking at the 4800u clocks i just don't see how it could be done whist also adding in a RDNA GPU part. Given that TDP is based on base clocks, going from 1.8 ghz to 3.6 is a tall order. I could be wrong of course.
Once AMDs Van Gogh chip is out in the wild we will have the information to make more accurate predictions given that it's most likely based on the Lockhart chip but downclocked for portable devices.
D
Deleted member 11852
Guest
You can still dream!
D
Deleted member 11852
Guest
Confirmed by TSMC earlier this month, testing 3nm Q1 2021, mass production by H2 2022.3nm is years out. Are you familiar enough with the characteristics of the node and Zen to write a portable off based on a 13% upclock?
- Status
