• Hey Guest. Check out your NeoGAF Wrapped 2025 results here!

Xbox Series X Memory Setup Explained in 3 Tweets

It will be very interesting to see if the split pool approach ends up hurting the XsX more than we would like

Either way it's a legit beast, but this seems to be an interesting design decision
 
Last edited:
Love everything about Series X design. This is another smart way to feed its powerful GPU.

Everything is laid out without cloudy half truths. Phil team have surprised even the most hard core doubters, no doubt
 
It will be very interesting to see if the split pool approach ends up hurting the XsX more than we would like

Either way it's a legit beast, but this seems to be an interesting design decision
I don't think it will make much of a difference, the "slow" memory is not that slow, and it's mostly meant to be used by the CPU (and there is not that much available for it in games).
 
Oh thats interesting I thought it was the 1gb chips which were the slower bandwidth.
 
PS3 was split pool, xbone was also split pool. Both designs were not looked upon favorably. In practice, it seems like MS has put in the effort to wallpaper over the hole so to speak (ie, developers may not run into issues).

honestly it is the one oddity that I found with the MS hardware. The rest seems pretty bulletproof and it is on Sony to justify the speedier SSD.
 
PS3 was split pool, xbone was also split pool. Both designs were not looked upon favorably. In practice, it seems like MS has put in the effort to wallpaper over the hole so to speak (ie, developers may not run into issues).

honestly it is the one oddity that I found with the MS hardware. The rest seems pretty bulletproof and it is on Sony to justify the speedier SSD.
This time is different. This design is smart. This is why unbiased core gamers are tipping their hats off Series X engineering.
 
This time is different. This design is smart. This is why unbiased core gamers are tipping their hats off Series X engineering.

Yea... just like PS4 I expect PS5 to start bitch slapping 2x powerful xbox from Feb to March 2021 onward with exclusives.

Remember xbox 1 was suppse to be infinitely powerful ...xbox1 plus cloud power.....while weak ps4 would suffer for the rest of eternity ...
 
Last edited:
It's going to be interesting to see next gen play out. I'm excited to see new games from 1st party and 3rd party

I also want to see what developers can do with the fastest SSD this year in the PS5 as it helps free up memory :)
 
PS3 was split pool, xbone was also split pool.

Neither the Xbone or XSX have split memory pools

The PS3 had 2 separate memory pools of 256Mb (Xbox 360 had 512 unified) which could only be accessed by one part of the system (CPU or GPU) where as the Xbone and XSX have single memory pools accessible to the entire system.

What Microsoft has essentially done is have a smaller segment of slower memory to save cost while keeping it accessible to the entire system.
 
Oh thats interesting I thought it was the 1gb chips which were the slower bandwidth.
That's my understanding too. I need to double check if i can find the original information now.

Edit:
That's the source of my confusion. But with the graphics it makes sense. It's better than my original understanding.
10 GB @ 560 GB/s, 6GB @ 336 GB/s
 
Last edited:
Oh thats interesting I thought it was the 1gb chips which were the slower bandwidth.
That's my understanding too. I need to double check if i can find the original information now.

Edit:
That's the source of my confusion. But with the graphics it makes sense. It's better than my original understanding.
10 x 1GB is the fast memory... that means half of the 2GB modules.
6 x 1GB is the slow memory... the other half of the 2GB modules.

MS choose that why it is cheaper... to make the full RAM to be 560GB/s for a 320bits they needed to use all 10 modules being 2GB.
That means 20GB instead 16GB.
 
Last edited:
This technical stuff goes right over my head.

I'll wait for this info in meme format.

pZgE6o4.png
 
Last edited:
10 x 1GB is the fast memory... that means half of the 2GB modules.
6 x 1GB is the slow memory... the other half of the 2GB modules.

MS choose that why it is cheaper... to make the full RAM to be 560GB/s for a 320bits they needed to use all 10 modules being 2GB.
That means 20GB instead 16GB.

Yes, there 6 2gb chips and 4 1gb chips. Those tweets explain it perfectly.
 
PS3 was split pool, xbone was also split pool. Both designs were not looked upon favorably. In practice, it seems like MS has put in the effort to wallpaper over the hole so to speak (ie, developers may not run into issues).

honestly it is the one oddity that I found with the MS hardware. The rest seems pretty bulletproof and it is on Sony to justify the speedier SSD.

Agreed but I think the RAM total for both consoles are disappointing. I see a lot people think it is 'fine' but I'm sure most would think the same thing with 4GB in 2013


While this pic would be apt in certain discussions, I wouldn't use it in a thread discussing RAM (both consoles share the same total amount where 38% of the allocation is slower on XSX)
 
But, how is this done:

But not all glasses in the #XSX are equal in size. Data that need high speeds, like textures, will use the first 1GB of each of the 10 glasses (10*56GB/s=560GB/s) while data that doesn't need high speed, like sound, will use the last 1GB of the bigger glasses (6*56GB/s=336GB/s)

How is it decided which "glass" the series x uses?
Do the developer need to specify this manually or is it done automatically?
 
But, how is this done:



How is it decided which "glass" the series x uses?
Do the developer need to specify this manually or is it done automatically?

There will likely be a flag when developers allocate memory that will indicate where they want it to go, if theirs free fast memory put it in there, if not, put it in the slow pool. This will be a manual process but its honestly not that much burden, it does get a bit tricker if you start to run out of fast memory but im sure they will work something out there.
 
I thought ram is the weakest part about these consoles. Does anyone know if 10gigs is enough for gfx ?

Xbox has 25% advantage here and this is most significant. Like last time it was with PS5 GDDR5 where the difference was massive to ddr3. Not the tflops. Even he said it is one of his tweets
 
So much fuss about the XBX RAM, while in reality, out of those 6 slower GB only 3.5 will be actually available for the games, which undoubtedly will be reserved for the CPU tasks, for which the bandwidth is more than enough, and probably some part of the faster 10GB will have to be used for the CPU as well, because I can't imagine next-gen titles to be able to fit within just 3.5GB...

In a similar way, PS5 will have ~13GB 448GB/s RAM for the games. And I actually think it would be better if the consoles have the opposite solutions given the PS5 Smart Shift technology, where slower ram for the CPU and faster for GPU would make perfect sense, whereas for XBX constant/guaranteed performance a fixed bandwidth RAM would also make more sense. But I guess MS did their calculations/tests and turned out they really need that bandwidth to feed that 12TF/52CU GPU.

And speaking of RAM, I wonder what happened that XBX uses only 2.5GB for the OS? They decided to stick with 1080p UI in X1X because the OS would take 4GB instead of 3, and now we will have 4K UI within just 2.5GB? Wonder what features have been removed in order to save up so much space?
 
But, how is this done:



How is it decided which "glass" the series x uses?
Do the developer need to specify this manually or is it done automatically?
Within the address space you allocate, the engine will determine 2 heaps (for ease of discussion) and this will be GPU and CPU. From here you just request against the mem pointer you need and the Mem Controller takes the request and flags back when it is ready for the requester.

CPU will get priority so this will still become a "segment" of game engines to manage requests as when the CPU or 6GB pool is being requested it will stall the GPU Block (faster ram) this is what I was talking about in my PS5 video in regards to contention.

PS5 will have a similar issue in that CPU will take Bandwidth from GPu as these are a single bus/same speed but the GPU scrubbers are a means for the PS5 to reduce these Mem requests or minimise them, i.e. reduce the bandwidth requirements. All this is very interesting for me as so many features between the system are different, looking forward to analyzing the launch games.
 
Yes, there 6 2gb chips and 4 1gb chips. Those tweets explain it perfectly.
Except DF said it has been done due to memory crosstalk and presumably in that case 20GB would require totally different PCB (or possible die) layout. I am not really sure, but the memory si pretty close to each other and die is really small, so that maybe the problem. But cost cuts is definitely issue in here.
 
Within the address space you allocate, the engine will determine 2 heaps (for ease of discussion) and this will be GPU and CPU. From here you just request against the mem pointer you need and the Mem Controller takes the request and flags back when it is ready for the requester.

CPU will get priority so this will still become a "segment" of game engines to manage requests as when the CPU or 6GB pool is being requested it will stall the GPU Block (faster ram) this is what I was talking about in my PS5 video in regards to contention.

PS5 will have a similar issue in that CPU will take Bandwidth from GPu as these are a single bus/same speed but the GPU scrubbers are a means for the PS5 to reduce these Mem requests or minimise them, i.e. reduce the bandwidth requirements. All this is very interesting for me as so many features between the system are different, looking forward to analyzing the launch games.
If you were working on a project, which consoles memory setup would you prefer for the game, even if by a bit? No need to answer if you feel your blast shield cannot weather rockets from one side or the other. :messenger_halo:
 
While this pic would be apt in certain discussions, I wouldn't use it in a thread discussing RAM (both consoles share the same total amount where 38% of the allocation is slower on XSX)
As long as we don't know how much RAM will be dedicated to the OS in case of the PS5, we can't really make the comparison. If they would be using the same amount of RAM (3.5GB) for the OS, the XSX would come out on top due to having more of the faster RAM compared to the slower RAM.
 
And speaking of RAM, I wonder what happened that XBX uses only 2.5GB for the OS? They decided to stick with 1080p UI in X1X because the OS would take 4GB instead of 3, and now we will have 4K UI within just 2.5GB? Wonder what features have been removed in order to save up so much space?

I think that might be down to them using the new Windows 10X Core OS. But if it really is 1080p again then that is shit.
 
Last edited:
PS3 was split pool, xbone was also split pool. Both designs were not looked upon favorably. In practice, it seems like MS has put in the effort to wallpaper over the hole so to speak (ie, developers may not run into issues).

honestly it is the one oddity that I found with the MS hardware. The rest seems pretty bulletproof and it is on Sony to justify the speedier SSD.

This is a unified memory pool; the two pools of fast/"slow" GDDR6 are treated as a singular pool of memory by the OS and to developers, and data assets don't need to be copied from the two pools for coherency purposes. And the memory pool is of the same technology.

Actual split pools are like the GDDR5/DDR3 in PS4 Pro, or DDR3/ESRAM in XBO, or the type of memory pools you had in MANY old-school gaming consoles. Different amounts of physical memory would be given to certain processors and only those processors, with their own bus interconnect to those processors exclusively. The memories were also usually of different types. However, with older game consoles you didn't have to copy data assets between the different pools for coherency like you do on PC with DDR system memory and GDDR graphics memory. Usually because on many older systems, programmers had actual low-level access, and usually worked with those systems in assembly as they didn't really have an OS to speak of. This even extends to systems like the PS2 and Gamecube (Xbox and Dreamcast had OSes tho, with Dreamcast's being voluntary use).

Even with more modern game consoles, despite having physically split memory pools, the OS usually handles a lot of the coherency/sync of data assets between them, so data assets don't need to have copies between the different pools nearly as much as on a PC. So the different pools are treated a lot like a virtually homogeneous pool of memory to developers, though they still have to manage what data goes where, and other small oddities, because the memory are of different technology types with their own specific ways of handling data operations (plus on a system like PS3 and XBO, the pools are physically split, with their own specific bus interconnect to either the APU (XBO), or CPU and GPU (PS3)).

XSX's approach is closer to more modern systems but the entire pool is the same GDDR6 memory technology. It also all (technically) physically connects to the APU on the same bus. Compared to actual split pools of the past and even split pools like in PS3 and XBO, developers don't have to deal with specific chunks of physical memory exclusively locked to different processors in the system, or micromanage flow of data across different memory technology structures. All developers need to do with XSX is prioritize what type of data goes into which partition of GDDR6 based simply on the type of data and how wide & fast they think it needs to be fed to components of the APU.

Within the address space you allocate, the engine will determine 2 heaps (for ease of discussion) and this will be GPU and CPU. From here you just request against the mem pointer you need and the Mem Controller takes the request and flags back when it is ready for the requester.

CPU will get priority so this will still become a "segment" of game engines to manage requests as when the CPU or 6GB pool is being requested it will stall the GPU Block (faster ram) this is what I was talking about in my PS5 video in regards to contention.

PS5 will have a similar issue in that CPU will take Bandwidth from GPu as these are a single bus/same speed but the GPU scrubbers are a means for the PS5 to reduce these Mem requests or minimise them, i.e. reduce the bandwidth requirements. All this is very interesting for me as so many features between the system are different, looking forward to analyzing the launch games.

Are you saying the XSX can't access both pools simultaneously? Or am I interpreting this wrong? If so that would sound like a bit of a design flaw and something MS would have worked on addressing IMO.

Or at least tried to, anyway.
 
Last edited:
Love everything about Series X design. This is another smart way to feed its powerful GPU.

Everything is laid out without cloudy half truths. Phil team have surprised even the most hard core doubters, no doubt

Yep, MS really have done an incredible job with XSX. I believe it's not possible to put more power and efficiency into a Console which cannot cost too much. They are giving us the maximum in terms of hardware. I still can't believe. XSX has exceeded my expectations, for the first time in all the generations.
 
This is a unified memory pool; the two pools of fast/"slow" GDDR6 are treated as a singular pool of memory by the OS and to developers, and data assets don't need to be copied from the two pools for coherency purposes. And the memory pool is of the same technology.

Actual split pools are like the GDDR5/DDR3 in PS4 Pro, or DDR3/ESRAM in XBO, or the type of memory pools you had in MANY old-school gaming consoles. Different amounts of physical memory would be given to certain processors and only those processors, with their own bus interconnect to those processors exclusively. The memories were also usually of different types. However, with older game consoles you didn't have to copy data assets between the different pools for coherency like you do on PC with DDR system memory and GDDR graphics memory. Usually because on many older systems, programmers had actual low-level access, and usually worked with those systems in assembly as they didn't really have an OS to speak of. This even extends to systems like the PS2 and Gamecube (Xbox and Dreamcast had OSes tho, with Dreamcast's being voluntary use).

Even with more modern game consoles, despite having physically split memory pools, the OS usually handles a lot of the coherency/sync of data assets between them, so data assets don't need to have copies between the different pools nearly as much as on a PC. So the different pools are treated a lot like a virtually homogeneous pool of memory to developers, though they still have to manage what data goes where, and other small oddities, because the memory are of different technology types with their own specific ways of handling data operations (plus on a system like PS3 and XBO, the pools are physically split, with their own specific bus interconnect to either the APU (XBO), or CPU and GPU (PS3)).

XSX's approach is closer to more modern systems but the entire pool is the same GDDR6 memory technology. It also all (technically) physically connects to the APU on the same bus. Compared to actual split pools of the past and even split pools like in PS3 and XBO, developers don't have to deal with specific chunks of physical memory exclusively locked to different processors in the system, or micromanage flow of data across different memory technology structures. All developers need to do with XSX is prioritize what type of data goes into which partition of GDDR6 based simply on the type of data and how wide & fast they think it needs to be fed to components of the APU.



Are you saying the XSX can't access both pools simultaneously? Or am I interpreting this wrong? If so that would sound like a serious system flaw and something MS would have worked on addressing IMO.
Yeah, poor choice of words on my part. The memory is of heterogenous performance, which is odd. It sounds like it's only 1 additional step of setting a priority for use but it's a step nonetheless.
 
I thought ram is the weakest part about these consoles. Does anyone know if 10gigs is enough for gfx ?

Xbox has 25% advantage here and this is most significant. Like last time it was with PS5 GDDR5 where the difference was massive to ddr3. Not the tflops. Even he said it is one of his tweets
The 2080 Super has 8. So 10 is enough here.
 
Yeah, poor choice of words on my part. The memory is of heterogenous performance, which is odd. It sounds like it's only 1 additional step of setting a priority for use but it's a step nonetheless.

True, it's an extra step, but devs had had to deal with far more complicated RAM setups in the past and worked around them. XSX's setup is super-simple by comparison to those, when you look at the scope of console history.

Honestly don't see it being an issue for next-gen development on the platform.
 
The 2080 Super has 8. So 10 is enough here.

Well, I have also seen videos where not more than 7gb is used on PC Vram. But PS4 was using 5.5gb for games and this time 13.5gb does not seem a big upgrade. It is just little over 2 times and little over 3times in terms of speed of Vram.
 
Top Bottom