Game Stack - Xbox Secret Sauce 😉

Let’s have a great hardware and software discussion here!

Xbox velocity architecture : Sampler feedback streaming

Saves VRAM memory by having a sampler feedback system in place.

It saves upto 67% of memory.

This results in upto 3x multiplier of memory for texture memory.

SFS allows bit by bit streaming of textures - more efficient from by LOD level streaming

SFS works on a hardware block of Xbox Series SoC (GPU)

The hardware gathers mip region usage data during scene rendering

This helps in loading LODs bit by bit

Latency :

The traditional hard drive takes 22 seconds to load the textures into the scene ( ~ 2 GB )

XVA and SFS does this in mere 0.19 seconds. That implies - in a 12 frames time for a 60 fps game around 2 GB of data can be loaded.

This is what ’ In snap of finger ’ means



SFS seems to be the real game changer, much more than VRS, which I intend as the next level of DRS (instead of dropping resolution, it drops rendering of less important areas of the screen), but it’s not 2/3x better than the old techniques, it’s 10/20/30% better depending on the engine/game. These two techniques in conjunction with DirectStorage will extend the life of Series S for an entire gen, considering that as of now, developers are not using these things at all.

They are also HW features on AMD/NVIDIA cards, not limited to Xbox and this thing will facilitate the implementation, bar Sony contracts, lmao.


Given how well the SSD on series consoles are already stacking to Ps5 with SFS it does have a shot of having the edge even on IO when SFS is used.

Another remarkably interesting thing to notice: Back when Ps5 was described Cerny talked about some scenarios that their IO solution would need. On the SFS demo all those scenarios described by Cerny are already shown fully operational here.

Given that same demo was done even before SX and SS were announced, while Cerny’s talk was more about hypothetical scenarios enabled by the super fast SSD, it’s a strong evidence on how it was indeed a central point of the Series consoles, while I’m not sure that the 2019 Ps5 version would have it for example.


Keep in mind that VRS can offer significant performance gains, but they may come with image degradation (like reducing the resolution does for example). And that the 10-30% perf increases are early implementations, focusing on not having any discernable IQ degradation (For example the one used by Coalition was developed by a single person in a couple of days).

There’s tons of potential for even better results, also as often happen in games, developers may find other uses for the hardware that enable other scenarios (Checkerboard rendering for example, was created because Ubisoft noticed a hardware used for AA on GCN had the ability to solve the checkerboard pattern with temporal data.

VRS is already being used for reducing the amount of rays in some implementations for example, which is crucial on consoles where seemingly the hardware to perform rays is more limited compared to high end pcs.

But overall it’s super exciting to have all those things on all fronts. I think tied together is what will make this generation really shine


I thought VRS was lowering the PQ/textures or w/e in parts where people aren’t actually looking to save performance so yeah it is lower PQ but it’s in parts of the screen you aren’t or shouldn’t be looking at.

But it allows for increased performance and/or resolution so overall it increases PQ.

1 Like

VRS should be used to lower the rendering resolution and therefore GPU load in parts of the picture where detail is low. If detail is low in a part of a frame (or better there is low frequency) you can’t notice lowered pixel shader resolution. Like the sky or blurred parts of the picture. It makes no sense to run your expensive pixel shader at full resolution and then apply motion blur over the result.

Devs have to prepare some stuff to work this out with Tier 2 VRS. Gears uses a edge detection filter, which basically means a filter to detect low frequencies where rendering in lower resolution doesn’t degrade picture quality.

Texture resolution should not be affected if VRS is implemented accurately (mipmapping thresholds have to be adjusted). VRS only lowers the rate of the pixel shader, so geometry is full resolution.


Yeah this is pretty much what I was saying, the person I was replying to I thought was saying overall PQ would be worse not just the parts that don’t matter, which is why I was saying I thought it was bits like in the distance that would look worse but the tree right in front of you looks really good.

Like in a racing game you will most likely focus on the track and the things in front of you so do you really need 4K textures on the crowd or grass to the bottom left of the screen haha. So you lower the textures or w/e to the crowd or grass no one is looking at to give more budget to like the hood of your car or track texture detail right in front of you, that’s correct right? haha.

in a racing game you would probably render the car or cockpit in full res. Everything else (grass, trees, road) moves at 100+ km/h, so is blurry as hell. No need to render that at full res.

I don’t know what any of this means, but I’m glad you’re all so excited :smiley:

IN short: Only load and show what you need for a scene. Thus loading times and memory usage is a lot less than traditional rendering and loading :).

Like if you bring drinks to your guest? They shout to you what they want and you only carry what you need to the table, rather than unload your entire cabinet and put it on the table ;).

1 Like

Ahhh, thanks very much for the explanation and analogy, it’s very helpful! So this would mean what we see is the same but the system is under less strain? And since it isn’t moving as much stuff, it’s more like consistently moving small things rather than one big thing every now and then?

Less to load, thus faster loading and less system/video memory in use as not everything has to be there all the time. It also means more detail for scenes were you stream the assets as you go along.

1 Like

It’s simply a better version of checkboarding.

CB forces low picture quality on complete scene. It becomes visibly in motion and edges

VRS has low quality on specific portion of the scene. Not visibly in motion or edges becuase it’s simply not applied on those parts. Dark portions is where VRS is used most. That’s why VRS gain is only 10-30 % but checkerboard can boost 50% + performance

Yes! It’s all about the final presentation.

It does reduce the quality, but in places that doesn’t matter or that you can’t see anyway.

Like it can group pixels by color, by velocity and focus.

For example, if there’s dof applied you don’t need to run a very expensive calculation on all of them as they will get blurred anyway.

But if you add enough info this can be done to any intermediate buffer as well such as lighting or shadow buffers


Long term this is gonna benefit xbox a lot

1 Like

Its rather ironic that this presentation sounded exactly what cerny was describing regarding the PS5 SSD. 2gb in 0.19secs is about what cerny suggested the PS5 could do (4gb in 0.5secs)

It seems playstation and xbox arrive at the same place, they just get there in different ways.

The most exciting thing for me is not greater quality textures but it will free up memory for better and more lighting+particle effects and physics simulations. Enviromental detail is actually the thing which I think needs the least improvement for my personal visual tastes.

Give me rdr2 world detail but with real time dynamic lighting, advanced smoke, wind, fire, snow, fluid simulations and increased game world object numbers (flock of birds, explosions with loads of debris, tornados with loads of debris)

1 Like