Xbox Series S RAM options

[invictis]

The RAM set up in the Series S is pretty underwhelming, and I wonder if they could have uses other options that would have been around the same cost, but a better outcome.

As it stands, the Series S has 10gb of GDDR 6, split into two pools, 8gb @ 224 GB/s and 2gb @ 56 GB/s. I fail to understand how the Series S would suffer from the same signalling issues that made MS go with the split pool on the Series X, nor how 2gb would be enough to cover the OS, CPU and audio on the Series S.

Could it have been a option for the Series S to have ran with 12gb of GDDR 5 RAM @ 326 GB/s bandwidth, similar to what's found on the One X?
It would have given more bandwidth than the current set up, more RAM, and the GDDR 5 costs should have been no more expensive than the 10gb of GDDR6 RAM.

Is there a reason they couldn't, or didn't, go with something like that as an option?
I understand their desire to keep the Series X and X as the same architecture, but surely the RAM could be different.

 

[BRiT]

nor how 2gb would be enough to cover the OS, CPU and audio on the Series S.

Series X only uses 2.5 GB for OS reservations, scale a few things down for resolution such as video capture buffers and dashboard overlay while gaming and you're down to 2GB.

 

[BRiT]

Could it have been a option for the Series S to have ran with 12gb of GDDR 5 RAM @ 326 GB/s bandwidth, similar to what's found on the One X?
It would have given more bandwidth than the current set up, more RAM, and the GDDR 5 costs should have been no more expensive than the 10gb of GDDR6 RAM.

That most likely requires a larger SOC for all the pins for the 384-bit memory interface. Doesn't seem economical if that's the situation.

Its all about performance within specified price region. Anything further and it wouldn't be priced as low.

 

[Jay]

It all would have been driven by cost, short term to long. Memory chip costs, cost to have AMD use GDDR5 instead of same design as XSX (just less CU's), for such a budget entry system all the costs matter even more.

The architecture being GDDR6 and split speeds doesn't really factor into it in terms of software.
As long as the size was at least 10GB and speed at least as fast as the fast pool.

 

[Proelite]

That most likely requires a larger SOC for all the pins for the 384-bit memory interface. Doesn't seem economical if that's the situation.

Its all about performance within specified price region. Anything further and it wouldn't be priced as low.

12GB GDDR6 needs 192 bit bus.

I think the whole motivation between slower and faster pools GDDR6 in the series consoles is to not increase die size due for OS partitions.

I.w you need 12GB ram for games and 336Gb/s ram bandwidth. Also need 2GB OS.

Instead of doing 14GB on 224 bit bus which increases die size and be to 389GB/s. You can put 2GB for the OS in clamshell, with 56GB/s speed for the OS. So you end up with 12GB on 192bit bus, 336GB/s instead. Both having similar perf but one is cheaper due to smaller die size, assuming clamshell on the mobo doesn't add as much cost.

 

[invictis]

Series X only uses 2.5 GB for OS reservations, scale a few things down for resolution such as video capture buffers and dashboard overlay while gaming and you're down to 2GB.

So unlike the Series X, the S split is only for the OS and the CPU and audio both come off the main 8gb of Ram?
Makes sense.

 

[BRiT]

So unlike the Series X, the S split is only for the OS and the CPU and audio both come off the main 8gb of Ram?
Makes sense.

OS reserve is 2.5 GB on Series X and comes from the slower memory pool.
OS reserve is 2 GB on Series S and comes from the slower memory pool.

Games on the X have use of 10 GB from the faster memory pool and 3.5 GB from the slower pool.
Games on the S have the use of 8GB from the faster memory pool.

 

[BRiT]

12GB GDDR6 needs 192 bit bus.

Ah, I got tripped up when I read "similar to what's found on the One X", so thats where 384bit came from.

 

[Proelite]

Ah, I got tripped up when I read "similar to what's found on the One X", so thats where 384bit came from.

I also added some more ramblings.

 

[peceed]

Xbox Series S doesn't need more memory bandwidth, it is perfect regarding gpu performance and target resolution (1080p).
But Series S desperately needs more memory capacity, they should go for 12 GB.
8 GB for games will constrict game logic of whole generation.
Games like Cyberpunk will suffer the most.
Series S without 12GB is the whole reason why PS5 is better than Series X

 

[invictis]

Xbox Series S doesn't need more memory bandwidth, it is perfect regarding gpu performance and target resolution (1080p).
But Series S desperately needs more memory capacity, they should go for 12 GB.
8 GB for games will constrict game logic of whole generation.
Games like Cyberpunk will suffer the most.
Series S without 12GB is the whole reason why PS5 is better than Series X

Lets look at GPUs with a similar performance level to the Series S.
Radeon RX 5500. This is a 5.2tflop GPU that targets 1080P and comes with 4gb of GDDR6 RAM and a bandwidth of 224gbs.
Radeon RX 5300. This is a 4.6tflop GPU that targets 1080P and it comes with 3gb of GDDR6 RAM and a bandwidth of 168gbs.
GeForce GTX 1060. This is a 4.4tflop GPU that comes with either 3gb or 6gb of GDDR5 RAM and a bandwidth of 192gbs.
GeForce GTX 1650 Super. This is a 4.4tflop GPU that comes with 4gb of GDDR6 RAM and a bandwidth of 192gbs.

So the Series S is 4tflops with 8gb of GDDR RAM shared between the GPU and CPU, and a bandwidth of 224gbs.
The Series X had 3.5gb of its RAM used for CPU, so one would think that the CPU requirement would be less than that on the Series S, so maybe 2.5gb. So this leaves around 5.5gb of RAM left over for the GPU.
So the average VRAM on both Nvidia and AMD cards for GPUs with more power than the Series S was between 3 and 4 gigs of VRAM.

Looking at this, it would tend to say that the Series S RAM shouldn't be too low.

 

[zed]

so one would think that the CPU requirement would be less than that on the Series S

Why? The CPU work would be about the same, theres lower GPU memory needed if your textures/buffers are smaller but the CPU stuff will be pretty constant between the 2.

Just picking the biggest recent game on PC, cyberpunk. minimum specs (not recommended, which will be higher)

• GTX 780 3GB (or RX 470 4GB)
• 8GB RAM.
• 3GB VRAM.

Im guessing most other recent PC games are similar 8GB memory & ~4GB VRAM. OK the OS for xbox should be leaner than windows, but I think they would of been better served with an extra 2GB

 

[peceed]

Lets look at GPUs with a similar performance level to the Series S.

All you have proven is that bandwidth is enough for 1080p.
What about capacity? The simplest model shares memory pool into resolution dependent "video" part (buffers, textures) and resolution independent part "logic". "Logic" has to be the same on two models as long as you want to have the same gameplay and fidelity.
"Video" size part is proportional to pixel count (4x difference for full hd vs. 4k) so "logic" memory is, 6.17GB, "video" is 1.83GB for Series S and 7.33GB for Series X.
As you can see, "video" memory is quite a low, but that is not a point. The real problem is that your game can not have bigger "logic" than 6.17 GB. You can think that it is not a big issue, but "logic" contains a lot of data that can not be scaled down with resolution: geometry, materials, global illumination data,
cached data for objects having small granularity, game logic, music, acoustic data.
2GB of memory (for $15) is the difference between comfort development and balancing on the edge.

 

[Deleted member 13524]

I'm pretty sure that having the OS taking "only" 2GB of GDDR6 doesn't mean the iGPU is getting full access to the remaining 8GB in the SeriesS.
When loading up a modern DX12 game in the PC, the executable will easily take 4 to 8GB of system RAM, plus the dedicated >6GB VRAM in the GPU.

There's a bunch of non-OS game data that is just handled by the CPU (caching, AI, sound, etc.). The high-ish speed of the NVMe solution probably allows for less caching to be done, but it's not like devs can just assume the GPU has access to 8GB VRAM. It may be closer to half of that.

I remember reading a bunch of developer statements indicating that the SeriesS' lower RAM amount is one of their biggest hurdles this generation.

 

[iroboto]

I'm pretty sure that having the OS taking "only" 2GB of GDDR6 doesn't mean the iGPU is getting full access to the remaining 8GB in the SeriesS.
When loading up a modern DX12 game in the PC, the executable will easily take 4 to 8GB of system RAM, plus the dedicated >6GB VRAM in the GPU.

There's a bunch of non-OS game data that is just handled by the CPU (caching, AI, sound, etc.). The high-ish speed of the NVMe solution probably allows for less caching to be done, but it's not like devs can just assume the GPU has access to 8GB VRAM. It may be closer to half of that.

I remember reading a bunch of developer statements indicating that the SeriesS' lower RAM amount is one of their biggest hurdles this generation.

in some ways, that may actually be comforting to Series S owners. If memory capacity is the issue, at least we know there are solutions for it coming. If compute and front end is the issue, it's going to be much harder to work around that as we assume those are already near optimum.

 

[Seanspeed]

All you have proven is that bandwidth is enough for 1080p.
What about capacity? The simplest model shares memory pool into resolution dependent "video" part (buffers, textures) and resolution independent part "logic". "Logic" has to be the same on two models as long as you want to have the same gameplay and fidelity.
"Video" size part is proportional to pixel count (4x difference for full hd vs. 4k) so "logic" memory is, 6.17GB, "video" is 1.83GB for Series S and 7.33GB for Series X.
As you can see, "video" memory is quite a low, but that is not a point. The real problem is that your game can not have bigger "logic" than 6.17 GB. You can think that it is not a big issue, but "logic" contains a lot of data that can not be scaled down with resolution: geometry, materials, global illumination data,
cached data for objects having small granularity, game logic, music, acoustic data.
2GB of memory (for $15) is the difference between comfort development and balancing on the edge.

I agree the capacity is a big problem, but I also dont think bandwidth is necessarily 'fine', either. It may have been alright for XB1/PS4 gen titles and hardware, but if Series S is supposed to be running the same performance targets as Series X, I think the bandwidth will start getting in the way with true next gen titles even at 1080p. Because 1080p isn't some locked demand. A next gen title at 1080p is still gonna be more demanding than a last gen game at 1080p.

And as for just having 2GB more, I agree that would have made a difference, but it also would have required a 192-bit bus rather than the 128-bit one. Which would have knocked out both birds with one stone - capacity and bandwidth. I'd have rather seen the system at like $320-330 with these upgrades.

As is, the Series S will get by, but it's not gonna be the system many originally envisioned it to be where a developer could just lower the resolution to 1080p and call it a day. Many games will clearly need extra work in downgrading the graphics to make it work. Which I cant imagine many are happy about given how much else they always have on their plate and with everybody demanding like two or three different performance profiles for each platform. And I do expect the compromises for Series S will get harsher the further into the generation we get.

 

[Dampf]

It will be fine, due to techniques like Sampler Feedback and DirectStorage.

Lets look at GPUs with a similar performance level to the Series S.
Radeon RX 5500. This is a 5.2tflop GPU that targets 1080P and comes with 4gb of GDDR6 RAM and a bandwidth of 224gbs.
Radeon RX 5300. This is a 4.6tflop GPU that targets 1080P and it comes with 3gb of GDDR6 RAM and a bandwidth of 168gbs.
GeForce GTX 1060. This is a 4.4tflop GPU that comes with either 3gb or 6gb of GDDR5 RAM and a bandwidth of 192gbs.
GeForce GTX 1650 Super. This is a 4.4tflop GPU that comes with 4gb of GDDR6 RAM and a bandwidth of 192gbs.

So the Series S is 4tflops with 8gb of GDDR RAM shared between the GPU and CPU, and a bandwidth of 224gbs.
The Series X had 3.5gb of its RAM used for CPU, so one would think that the CPU requirement would be less than that on the Series S, so maybe 2.5gb. So this leaves around 5.5gb of RAM left over for the GPU.
So the average VRAM on both Nvidia and AMD cards for GPUs with more power than the Series S was between 3 and 4 gigs of VRAM.

Looking at this, it would tend to say that the Series S RAM shouldn't be too low.

Mostly right, remember cards like the RX 5500/1060 do not have advanced DX12U featuresets nor HW-accelerated Raytracing, so they should not be compared as the Series S will be leagues ahead of these cards in true next generation games. Series S should have around 4-5 GB video memory too (CPU RAM requirements won't change between XSX and XSS because they have the exact same CPU), but Sampler Feedback Streaming should make that behave more like 10 GB.

 

[Ronaldo8]

It will be fine, due to techniques like Sampler Feedback and DirectStorage.



Mostly right, remember cards like the RX 5500/1060 do not have advanced DX12U featuresets nor HW-accelerated Raytracing, so they should not be compared as the Series S will be leagues ahead of these cards in true next generation games. Series S should have around 4-5 GB video memory too (CPU RAM requirements won't change between XSX and XSS because they have the exact same CPU), but Sampler Feedback Streaming should make that behave more like 10 GB.

Yep. Sampler feedback acts as an effective RAM multiplier. More importantly, however, is that one's texture budget is ultimately limited by what one's GPU can actually draw & shade within a given frametime.

 

[Seanspeed]

I see a lot of people thinking SFS is gonna be 'the answer' to XSS's low memory size, but this rationale ignores that what's important is the relative difference to the XSX.

It'd be one thing if SFS only worked on XSS as a way to overcome its limitations, but developers will be utilizing SFS on XSX as well in order to push their ambitions and make the most of that 16GB. It's going to be critical for devs to do so because 8GB to 16GB of RAM is a historically *pathetic* jump in capacity for a new generation. We're used to getting more like an 8x or even 16x multiplier in RAM, and now we've only got 2x. So the SSD + sampler feedback streaming is going to be key for devs to push proper next gen experiences.

And if you've got devs pushing what the XSX can do on this front, the XSS falls into the same problem of having a relative lackluster RAM setup.

 

[Silent_Buddha]

I see a lot of people thinking SFS is gonna be 'the answer' to XSS's low memory size, but this rationale ignores that what's important is the relative difference to the XSX.

It'd be one thing if SFS only worked on XSS as a way to overcome its limitations, but developers will be utilizing SFS on XSX as well in order to push their ambitions and make the most of that 16GB. It's going to be critical for devs to do so because 8GB to 16GB of RAM is a historically *pathetic* jump in capacity for a new generation. We're used to getting more like an 8x or even 16x multiplier in RAM, and now we've only got 2x. So the SSD + sampler feedback streaming is going to be key for devs to push proper next gen experiences.

And if you've got devs pushing what the XSX can do on this front, the XSS falls into the same problem of having a relative lackluster RAM setup.

It depends on the title. If we're talking multiplatform titles, where cross platform parity is the general rule of thumb, then it'll be far more important for the XBS-S than for the XBS-X. While there are some developers that will push the limits on the cross platform title, they are pretty rare. Basically, if SFS plays out and is easy enough for a cross platform developer to use, then the PS5 potentially becomes the limiting factor WRT memory size. However, that assume that a cross platform developer bothers to use SFS, something we just don't know. Without that then an argument could be made that XBS-S is the limiting console. Of course, this also doesn't take into account whether or not something similar to SFS can be used on the PS5 and/or whether it's faster SSD performance will allow it to swap things out fast enough that it isn't hampered by a lack of SFS.

That, of course, also ignores the resolution differences between what the XBS-S is expected to render compared to the other consoles which has an impact on memory requirements as well.

Regards,
SB