Xbox Series X: A Closer Look at the Technology Powering the Next Generation

Axel Stone

Member
Jan 10, 2020
810
How I see it;
We know how a CPU and GPU affect a game, and we know how solid state memory will affect game design, but I don't know how the extra 5gbs of speed on the PS5 SSD will do much better for games compared to xsx
I thought this was quite a nice post to that effect in the PS5 tech thread:


Edit: the text in the link preview isn’t the post I mean.
 

nib95

Banned
Oct 28, 2017
12,967
Yeah, CPU has slower access in comparison, but GPU has still full bandwidth. And CPU blocking GPU access to some modules when accessing data is the same in both architectures, but GPU can still access RAM modules that CPU is not accessing in given moment.
Right, but I wasn't talking about GPU access to the ram that the CPU wasn't accessing, but GPU access to that ram when the CPU was accessing it.

My point being that the faster memory bandwidth on the 10GB of GPU ram on the XSX, is offset somewhat by the slower 6GB CPU ram (due to the GPU having to wait during CPU access) and the CPU ram bandwidth being slower. That and the XSX's CPU ram access being slower than the PS5's, also offsetting some of the XSX's GPU ram side bandwidth advantage.

Then my second point was about the difference if games do require or use more than 10GB on the GPU side.

I'm sure they'll both have lots of stuff that mitigates a lot of the usual CPU memory overheads, but at the same time I'd imagine next gen devs make far more use of the CPU too, especially on things like AI and Physics. The weak CPU was commonly seen as a major bottleneck of this current gen, and given the much more potent CPU's of the next gen systems, of course they'll be capitalised on.
 

M3rcy

Member
Oct 27, 2017
662
It isn't necessarily better or as clear cut as you're making it seem for two reasons.

1.) Because of the way ram works, every time the XSX's CPU has to access ram (which is going to be routinely), it does so at the slower 336 GB/s speed, during which the GPU essentially has to wait its turn before it can access the ram. So in that sense the GPU is beholden to the CPU memory bandwidth and the entire average bandwidth is going be less thanks to the slower 6GB. Comparatively the CPU in the PS4 will access its ram at the same 448 GB/s as the GPU, so irrespective of whether the GPU or CPU access the ram, it's the same speed for both, and that is higher than the XSX's CPU ram access bandwidth.

2.) Add to the above, we don't actually know how much video ram will be available to devs with the PS5. If the PS5 offers more for GPU use than the XSX (eg more than 10GB), which might be the case if theories from NXGamer and others hold any weight, then the PS5 has yet another bandwidth (and memory) advantage, as not only will its CPU have faster memory access, but it's GPU will have faster bandwidth over everything beyond 10GB too.
Again, the CPU isn't consuming nearly as much data. So having a slower bus to the data it will be using is not as impactful as you seem to believe. It will take more cycles for the CPU in the XBSX to get a given amount of data than it will for the PS5's CPU to get an equivalent amount of data. It will also take the PS5's GPU more cycles to get a given amount of data than it will take for the XBSX GPU to get an equivalent amount of data. Which do you think will be more detrimental to the system performance?

As for your second point, there's no reason why developers would ever be accessing performance critical data from the slow pool. They will make sure this never happens when developing/optimizing the game. What the PS5 RAM allocation is won't effect how memory is used on the XBSX one iota.

I've said elsewhere that PS5 may very well have an effective memory capacity advantage due to the faster SSD (in addition to the generic "faster loads"). That's where it's advantage lies when it comes to memory setup.
 

Landy828

Member
Oct 26, 2017
6,508
Clemson, SC
Now that we've seen all the specs, I'm pretty sure I'll be buying an XSX day one.
I'll get a PS5 sometime in 2021.
That has been my way of doing things for the last couple generations.

Microsoft's hardware for all my BC games and Sony's hardware for exclusives. I bought a PS4 when the slim came out, and got all of Sony's exclusives.

I'll buy the series X when it comes out in order to have all my backwards compatible stuff, and to play 3rd party stuff. I'll buy the PS5 when Sony has a bunch of great exclusives under their belt.

I'm REALLY, REALLY excited for the new Xbox. I'm really impressed with Microsoft focusing "across the board". They went with the higher end GPU, CPU, and they focused on a fast SSD interface. It's a really well thought out package. I can't wait wait to get my hands on it.
 

bsigg

Member
Oct 25, 2017
10,808
What do you prefer at launch? Xbox Series X at $499 with one year of game pass or $399?
Obviously $399 at launch for Series X would be fantastic but I wouldn't expect less than $499.

Also I think they'll only toss in like a week trial to GP/GPU. Their goal is to get you to subscribe ASAP and then give you a reason to hang around.
 

DocH1X1

Member
Apr 16, 2019
891
copy paste from a Stadia thread:
CPU
PlayStation 5
: Custom Zen 2 CPU with eight Cores at 3.5 GHz (Variable Frequency)

Xbox Series X: Custom Zen 2 CPU with eight Cores at 3.8 GHz (3.66 GHz with Simultaneous Multithreading)

Stadia: Custom 2.7GHz hyper-threaded x86 CPU with AVX2 SIMD and 9.5MB L2+L3 cache


GPU
PlayStation 5
: Custom RDNA 2, 10.28 TFLOPs, 36 CUs at 2.23GHz (variable frequency)

Xbox Series X: Custom RDNA 2, 12.15 TFLOPs, 52 CUs at 1.825 GHz

Stadia: Custom AMD GPU, 10.7 TFLOPs, 56 CUs guestimated to run at 1.495GHz


Memory
PlayStation 5
: 16GB GDDR6 with a 256mb bus | 448GB/s bandwidth

Xbox Series X: 16GB GDDR6 with a 320mb bus | 10GB at 560 GB/s, 6GB at 336 GB/s bandwidth

Stadia: 16GB of HBM2 RAM | 484GB/s bandwidth
I genuinely didnt realize stadia had a more powerful gpu and ram than ps5.
 

idioteque

Member
Nov 8, 2017
207
Random thought, I'd love it if the XSX were to ship with Halo Infinite on the hard drive pre-installed. Would encourage people to give their ganepass trial a spin.

Appreciate it's probably impossible to do though.
 

Fredrik

Member
Oct 27, 2017
6,282
I genuinely didnt realize stadia had a more powerful gpu and ram than ps5.
Stadia is a beast ;)
Seriously though, the CPU is slower and the GPU is an older generation, but it’s not too shabby. And this is just the start, we still don’t know if or when Google will upgrade or how much more power they can get when using multiple Stadia blades. And we don’t know the SSD speed either which is so important nowdays at this place. There is more uncertainity regarding Stadia than how long it’ll take until Google will close it.

At this point XSX clearly takes the power crown though, and with that xCloud as well (I think, unless Geforce Now is already ahead?)
 

jelly

Member
Oct 26, 2017
13,104
Its about space also. Series X processor can run multiple Xbox One's at the same time.
If that's the case then would Series X run two Lockharts rather than having Lockhart blades if the thing is real that is then when XCloud matures they transition away from Xbox One to Lockhart/Series X virtualisation. Seems kinda smart and cost effective. You have one blade design that can run three consoles instead of three blades.
 

Samusaran

Member
Oct 27, 2017
6,173
Stumbled across this article looking a bit more deeply at the dynamic latency stuff on XSX:

This deserves it's own thread
 
Mar 20, 2018
31
I want straws in my XSX like the graphic.

But seriously, not to repeat what other's have said, but I bought into the One at launch, then waited for One X. I never bit on a PS4, but I will eventually with the PS5. I'll be hitting F5 for preorder on the Series X once it is available.
 

LordBlodgett

Member
Jan 10, 2020
374
I don't think Microsoft will develop a different chipset for Lockhart. If anything it will be binned Scarlett chips.
Nah, I completely disagree with that. First they are going to have pretty decent yields with their more conservative clocks, and yields will only increase over the lifetime of the console. Second the way that companies pay for chips (by the wafer) irregardless of the performance they designed means that a smaller chip would save a lot of money. That's not to mention they can go for smaller case, cheaper cooling solution, smaller and cheaper motherboard, etc.....
 

LordBlodgett

Member
Jan 10, 2020
374
I genuinely didnt realize stadia had a more powerful gpu and ram than ps5.
They don't. the efficiency gains from jumping from GCN (which Stadia runs on) to RDNA1 is pretty huge. not to mention both Sony and Microsoft will be using RDNA2, which I would guess has at least some efficiency gains over RDNA1. There is a reason why a Radeon 5700XT (9.75 TFLOPS on RDNA1) offers better performance in every game than the Radeon Vega 64 (12.66 TFLOPS on GCN) even though it is lower TFLOPS number. Architecture efficiency is a real thing. Honestly Stadia was kind of foolish to launch on GCN when it did. They must have gotten a killer deal on the APU's
 

DukeBlueBall

Member
Oct 27, 2017
3,945
Seattle, WA
I genuinely didnt realize stadia had a more powerful gpu and ram than ps5.
It doesn't.

If that's the case then would Series X run two Lockharts rather than having Lockhart blades if the thing is real that is then when XCloud matures they transition away from Xbox One to Lockhart/Series X virtualisation. Seems kinda smart and cost effective. You have one blade design that can run three consoles instead of three blades.
Xsx has the same CPU as Lockhart, so it can't run 2x Lockhart instances.

But a Lockhart blade probably can run 3 Xbox 1s instances.
 

Colbert

Banned
Oct 27, 2017
6,174
Germany
Xsx has the same CPU as Lockhart, so it can't run 2x Lockhart instances.

But a Lockhart blade probably can run 3 Xbox 1s instances.
a) MS said they can run 4 X1S cloud instances, your math would imply that it should be able to run 6 X1S instances. From where do you derive that information?

b) If a single XSX SOC can run 4 X1S instances from the cloud why would you make Lockhart based blades that can run less instances if we assume a 4 processor setup per blade? To deploy XSX blades is a more cost effective way to deploy xCloud capacity just because of important factors like occupied rack space per compute, performance per watt and scalability (increase of resolution for instance).

What you say makes no sense to me especially under the current base of information given by MS.
 
Last edited:

OneBadMutha

Member
Nov 2, 2017
4,917
Trying to understand why some people are trying so hard to point out this major bottleneck design flaw that has not been identified as a bottleneck by any reputable, established site. A thing which Goosen has explained. Theres no goofy Esram setup for devs to ignore. Outside of that one time, no Microsoft console has been bottlenecked on RAM. Seems illogical to add a beast of a GPU that will never be able to run as fast as it should because a much cheaper component was skimped. A lot of work to downplay.

Moving on...at this point doesn't matter what other plastic is doing. This one is a step beyond bad ass. What I'm curious about is the ML capabilities. We've seen some limited examples. It seems like this console was optimized for ML by being able to use a lot more of the less precise floating point operations. Am I off base in that? Anyone here have insight or guesses to how it may be used next gen? Is primarily used to upscale and add HDR or do we expect that's just the start?
 

Colbert

Banned
Oct 27, 2017
6,174
Germany
Trying to understand why some people are trying so hard to point out this major bottleneck design flaw that has not been identified as a bottleneck by any reputable, established site. A thing which Goosen has explained. Theres no goofy Esram setup for devs to ignore. Outside of that one time, no Microsoft console has been bottlenecked on RAM. Seems illogical to add a beast of a GPU that will never be able to run as fast as it should because a much cheaper component was skimped. A lot of work to downplay.

Moving on...at this point doesn't matter what other plastic is doing. This one is a step beyond bad ass. What I'm curious about is the ML capabilities. We've seen some limited examples. It seems like this console was optimized for ML by being able to use a lot more of the less precise floating point operations. Am I off base in that? Anyone here have insight or guesses to how it may be used next gen? Is primarily used to upscale and add HDR or do we expect that's just the start?
The XSX SOC is a dual purpose SOC (home console & cloud). I personally do not expect much of use of the int8 and int4 capabilities in the home console environment other than execute learned patterns but it will be used a lot in the cloud space and that is why MS put it in. That way the resources deployed for xCloud can also be leveraged for the growing need of ML tasks in the general cloud. This I think will give MS a big edge in economies of scales and cost mitigation for the SOC that will help to offer the home console at a better price.
 

gremlinz1982

Member
Aug 11, 2018
2,221
Trying to understand why some people are trying so hard to point out this major bottleneck design flaw that has not been identified as a bottleneck by any reputable, established site. A thing which Goosen has explained. Theres no goofy Esram setup for devs to ignore. Outside of that one time, no Microsoft console has been bottlenecked on RAM. Seems illogical to add a beast of a GPU that will never be able to run as fast as it should because a much cheaper component was skimped. A lot of work to downplay.

Moving on...at this point doesn't matter what other plastic is doing. This one is a step beyond bad ass. What I'm curious about is the ML capabilities. We've seen some limited examples. It seems like this console was optimized for ML by being able to use a lot more of the less precise floating point operations. Am I off base in that? Anyone here have insight or guesses to how it may be used next gen? Is primarily used to upscale and add HDR or do we expect that's just the start?
Well, Microsoft did state that they wanted silicon that they could use for gaming and enterprise. I wonder what all of that achieves when it comes to pricing.
 
Oct 27, 2017
3,914
Florida
The XSX SOC is a dual purpose SOC (home console & cloud). I personally do not expect much of use of the int8 and int4 capabilities in the home console environment other than execute learned patterns but it will be used a lot in the cloud space and that is why MS put it in. That way the resources deployed for xCloud can also be leveraged for the growing need of ML tasks in the general cloud. This I think will give MS a big edge in economies of scales and cost mitigation for the SOC that will help to offer the home console at a better price.
If they implement DirectML upscaling like DLSS it would be useful.
 

PianoBlack

Member
May 24, 2018
699
I'd love to read a deep dive on how the XSX SOC runs 4(!) concurrent XB1 games. Do they have to duplicate parts of the chip that they otherwise wouldn't? How are resources assigned and kept separate? SOC running with SMT without SMT? Does each game get assigned 2 cores/4 threads, or is code for all 4 games distributed across all cores arbitrarily?

And they must have additional RAM for the blades. 16GB / 4 is nowhere near enough RAM for each XB1 game.
 

Terbinator

Member
Oct 29, 2017
1,638
I'd love to read a deep dive on how the XSX SOC runs 4(!) concurrent XB1 games. Do they have to duplicate parts of the chip that they otherwise wouldn't? How are resources assigned and kept separate? SOC running with SMT without SMT? Does each game get assigned 2 cores/4 threads, or is code for all 4 games distributed across all cores arbitrarily?

And they must have additional RAM for the blades. 16GB / 4 is nowhere near enough RAM for each XB1 game.
Be down to a mix of VMs/instances and the SSD, i'd have thought in addition to usable RAM/CPU power.
 

Colbert

Banned
Oct 27, 2017
6,174
Germany
If they implement DirectML upscaling like DLSS it would be useful.
AMD has 2 technologies that are currently better than this. FidelityFX and AMD Resharpener.
We also heard from MS research that they experimenting with Asset compression to reduce them in size and still maintain the quality by 90%. For the latter one the learning will take place in the cloud just like it is with DLSS. Only the learned pattern is then executed on the client device, in this case the home console. Thus I expect the usage of the ML capabilities more on the backend side than on the front end side.
 

KKRT

Member
Oct 27, 2017
1,226
Right, but I wasn't talking about GPU access to the ram that the CPU wasn't accessing, but GPU access to that ram when the CPU was accessing it.
I'm not so sure that they limit itself to this degree, as they definitely need to use separate channels to handle different bus widths. It really depends how memory controller that handles both CPU and GPU works.
 
Oct 27, 2017
3,914
Florida
AMD has 2 technologies that are currently better than this. FidelityFX and AMD Resharpener.
We also heard from MS research that they experimenting with Asset compression to reduce them in size and still maintain the quality by 90%. For the latter one the learning will take place in the cloud just like it is with DLSS. Only the learned pattern is then executed on the client device, in this case the home console. Thus I expect the usage of the ML capabilities more on the backend side than on the front end side.
I'm not a huge fan of cloud dependency for graphical features
 

chipperrip

Member
Jan 29, 2019
171
I'd love to read a deep dive on how the XSX SOC runs 4(!) concurrent XB1 games. Do they have to duplicate parts of the chip that they otherwise wouldn't? How are resources assigned and kept separate? SOC running with SMT without SMT? Does each game get assigned 2 cores/4 threads, or is code for all 4 games distributed across all cores arbitrarily?

And they must have additional RAM for the blades. 16GB / 4 is nowhere near enough RAM for each XB1 game.


It's probably just saving an image of the game instance from ram to the drive. Sort of like an emulator save state, or the hibernate (not sleep) function on Windows.

When stored, the game is effectively not taking up any resources except storage.

Their limit of 4 games is likely just an arbitrary choice to avoid reserving too much storage space for this feature.

Edit: In this post I mistakenly thought PianoBlack was referring to Quick Resume.
 
Last edited:

Scently

Member
Oct 27, 2017
592
It's probably just saving an image of the game instance from ram to the drive. Sort of like an emulator save state, or the hibernate (not sleep) function on Windows.

When stored, the game is effectively not taking up any resources except storage.

Their limit of 4 games is likely just an arbitrary choice to avoid reserving too much storage space for this feature.
You are talking about something different from what that person wrote. The XSX SoC can run four X1S games simultaneously.