7 March 2026

Analyse This: The Rendering Cost of AI...

Unironically, I used AI for this...


The interest surrounding the rumours of the next gen consoles has been slowly boiling over the last few months. Now, with new confirmations from Xbox's newly-minted CEO, Asha Sharma we have a firmer grasp of what the next gen holds for one side of the equation. On the other side, Playstation 6 also has a lot of hardware specification leaks but one thing still confuses me: why all the memory?!

In this post, I'm going to look at the logic of having a lot of RAM in the consoles and why, to my mind, the leaked specs for the memory configurations just don't make sense...


As of 2023, 24+ Gbit modules were not on the roadmap... (TechPowerUp)



Spinning Around...


Xbox have come out swinging after the negative PR surrounding Phil Spencer's and Sarah Bond's exit from the division. Asha has given some small murmurings so far until a tweet yesterday which confirmed that Microsoft is going for a system which will be both a console and PC. This is a very interesting move and something which was being called for since back when the Xbox Series consoles were on their way. 

The sad thing is that this endeavour has obviously been in the works under the leadership of Phil and Sarah - and neither will see the credit for such a bold project.


Getting back to the discussion at hand, Tom over at Moore's Law is Dead is talking about how the higher specs of the next gen Xbox (the Magnus APU) will play to the strengths of that focus and I can't really deny that - especially with the included Neural Processing Unit (NPU) that will be able to run some AI models independently from the GPU compute infrastructre. However, Tom, himself, stated that it this is supporting up to 48 GB GDDR7 - which would require 32 Gbit / 4 GB modules. 

Now, I have been unable to find any mention of such a capacity for two of the three GDDR7 manufacturers and, as far as I can tell, only Micron has listed an updated roadmap that shows above 24 Gbit / 3 GB modules... and there have been no announcements of commencement of manufacturing of these products aside from Nvidia's announcement of a product which is expected to launch at the end of 2026. So, that leaves us with the possibility of 36 GB RAM as a maximum on the next gen Xbox console. 

Why 36 GB? Well, it's simply because no recent console has ever used the cutting edge memory tech. Playstation 4 was using 5.5 Gbps (Xbox One was even lower spec!) GDDR5 when a year prior, Nvidia graphics cards were already using 6 Gbps modules. The mid-gen refreshes of the One X and PS4 Pro were also lagging behind the already released GTX 9 and 10 series. Playstation 5 and Xbox Series were using 14 Gbps GDDR6 when two years earlier Nvidia cards were using the same memory in the RTX 20 series (with RDNA 1 cards following a year later in 2019). Even the Playstation 3 was using 1300 Mbps and the Xbox 360 was using 1400 Mbps GDDR3 when the 7800 GTX 512 was using 1600 Mbps a year earlier.

So, for people to assume that a memory density which was not officially announced that I can find, or referenced outside of a technical paper from JEDEC - nor from the big three (Micron, SK Hynix, and Samsung) will suddenly appear on the market for a mass-produced consumer good is pretty crazy! Hell, we've been hearing about 3 GB GDDR7 modules for years and we still don't have them on consumer GPUs! (Though it seems that's about to change...) 

That likely means we have at least a two year gap between spin-up and datacentre exclusivity before they appear on high-end, specialised consumer products (i.e. not consoles). There simply won't be the bulk production of the part, nor secondary supplier availability to supply a console release.


My appraisal of the potential memory configurations of the next gen consoles with a comparison for two currently available...


Written in the Stars...


Moving over to Playstation, Tom shared that Sony could have 30 or 40 GB RAM and, more recently, he provided the caveat that nothing is decided yet for both consoles and that they could have less. That's fair enough - and is often something which is left out of "leak" information. They're all too often provided with the certainty of an AI chatbot.

Here's where I begin to have a problem, though: most consoles in recent memory haven't used a clamshell design.

Seriously, let's look at: PS3, Xbox 360, PS4 Pro, Xbox One, Xbox One X and One S, PS5, Xbox Series X and S, and PS5 Pro, Steam Deck, (I believe all Aya Neos, but here's one), ROG Ally and Ally X, Switch and Switch 2.

The only modern console to feature a clamshell design was the base Playstation 4 - and that's likely an outcome of the late specification change to double the memory capacity - a la Randy Pitchford.

There are three reasons (that I'm aware of) why clamshell design isn't generally favoured and implemented:
  • Complexity of circuit board design
  • Complexity of device design
  • Cost to implement
Looking at the circuit boards of most modern consoles (the PS5 Pro and Series X pictured here), surface mounted devices are on the rear of memory module locations...


The first point is pretty self explanatory - you need to design the motherboard differently to allow for clamshell connections. This may require more layers to avoid signal issues, or even to allow the extra routing which must take place to account for the different spatial locations of the DRAM pins.

The device (i.e. the console) may also now have a more complex design - especially with regards to thermal management. Having GDDR7 chips on top of each other may result in a thermal trap which could lead to long-term reliability issues/higher failure rates. Thus, it's likely that more heatsink design must take place - and that could potentially affect the physical design and shape of the product!

These aspects cause a higher cost to both design and manufacture the circuit board - something which a console manufacturer will be keen to keep at a minimum! Additionally, purchasing more memory modules, in the middle or tail-end of a RAM shortage is not going to be favourable from a procurement standpoint, nor a pure cost standpoint. Fewer modules are better, in general! Along with this, there are additional costs related to the testing of the memory, once mounted on the board, along with having to address potential issues with cooling the chips on both sides of the board instead of just using one point of contact (either on the chip or underneath it).

All in all, these are not last minute additions to a project.

So, in my read of the situation, either Microsoft and Sony have already accounted for increased numbers of chips on a single side of the circuit board or they have already designed the board and cooling solutions to allow for clamshell arrangements of the memory modules...

The latter seems unlikely to me. The former? Well, let's move onto that.


What's the Point..?


Here's where I get into the logic of having a lot of RAM on the consoles because I just can't get past the "WHY?".

Why would you have such over-provision of memory?

Consoles are historically very price sensitive devices but even beyond that, mass produced, commercial devices are all very cost optimised - sometimes to their detriment. So, going on this logic, the amount of memory must be utilised for something and not present, just because...

A modern, graphics-heavy title is typically using 15 - 18 GB VRAM when using all modern features at 4K resolution... (TechPowerUp)



Games are typically using around 8 GB system RAM combined with 8-10 GB VRAM on a PC and approximately 12-13 GB of memory on the consoles. This puts even the maximum game requirements at 13 - 18 GB memory - even for a next gen title since even next gen consoles won't be rendering games at native 4K, they will be using upscaling from a lower resolution - such as 1080p (or lower).

Checking the recently released Resident Evil Requiem's VRAM usage on an RX 9060 XT 16 GB, at 1080p, Max quality settings with RT High - the game was averaging around 11 GB VRAM at native and when using upscaling at the balanced setting. There are other games which also use these advanced features which use less VRAM. E.g. Doom The Dark Ages and Assassin's Creed Shadows.

So, if a theoretical next gen console could get away with remaining on 16 GB VRAM or even 18 GB for a modest increase, why would Sony and Microsoft be looking at double that quantity?


The big explanation appears to be "AI". But, again, this is where I become skeptical.

"AI" doesn't just happen. It either needs to be contracted out to a cloud-based service - in which case games wouldn't be able to rely on such a feature because 1) latency and 2) connectivity issues - OR it needs to be run locally.

Local models are generally expensive - both from a computational standpoint but also from a memory and storage footprint aspect. Let's leave aside the storage issue for the time being because, in the grand scheme of things, it's irrelevant to this discussion. The memory footprint is a bigger deal, though.

Image generation is unlikely to be "useful" for use during games, so I think we can discount that (though I will include such a scenario in a moment) but AI models might be conceivably used to generate dynamic text or voice content.

LLMs have a variety of models available which can require anywhere from 2 GB to 12+ GB. Typically, larger models are more accurate but we don't expect a 4+ billion parameter model to be shoehorned into a console.... do we?! For voice generation, finding numbers was a bit more difficult but it seems that around 4-8 GB is a reasonable range. So, adding those together, you'd expect to be needing around another 16 - 18 GB of memory, on top of what's required for the OS and game.

So, yeah. I can see why people would be expecting 30 - 40 GB on both next gen consoles.

But there's another problem...


Performance numbers for a couple of games when running two types of AI in parallel...



Reality Strikes...


The problem is that "AI" isn't free. In addition to the memory, it needs both large amounts of compute as well as memory bandwidth - both of which are anathema to actually running a game on the hardware. Games, and specifically high-end effects like ray tracing, require a lot of bandwidth from the memory. To run an AI workload in parallel with the rendering workload, along with the frame draw calls and various simulation systems (e.g. collision detection, pathfinding, etc) will absolutely kill performance.

Even without running a game, the memory controller on the GPU was doing some serious work and the compute requirements were also really loading up the GPU core! These are not light workloads and they are not designed to be run concurrently with a real gaming workload.


Memory controller saturation and compute requirements for running even simplistic applications is overwhelming...


The informal testing I performed for this blogpost showed between 15 - 45 % fps loss when using a chatbot during games where the VRAM limit wasn't exceeded, depending on the specific game. Obviously, for image generation, it was quite easy to exceed that limit but for God of War and AC:Odyssey, I was able to keep within the VRAM capacity and still saw significant losses. 

As I said above, image generation won't be utilised during a game - I just can't see a useful application. However, you could envision this as a stand-in test for concurrent text generation and text-to-voice for use in a game. In that light, the performance loss is within the expected ballpark for such a nightmare scenario.

Unfortunately, I wasn't able to test any text-to-speech solutions as I didn't find any that could be set up easily (considering I'm not planning on using them)...

Looking at the supposed specs of the PS6 and Next Gen Xbox, I don't think they're that far beyond what we currently have - an RX 9070 XT with an extra 10% rendering performance (considering potential clock frequency and/or power limits that can and are usually applicable for the consoles) isn't going to magically make these costs disappear.

Yes, the Xbox supposedly has an NPU with 110 TOPs (Int 8?) but the 9060 XT has double that, and the 9070 XT has 3.5x that - but the NPU still needs to pull from the memory bandwidth at the same time as a game is trying to render.

Imagine the game is running and all of a sudden there are dips and drops in performance due to AI being activated. It's an untenable situation. The only logical thing for developers to do would be to assume that they can only count on the lower performance available in worst case scenarios. Thereby negating the true power of the GPU in the device. That doesn't sound logical to me! 


Conclusion...


Yes, this is testing on a low-end GPU but the principle still stands. AI will absolutely murder the available performance of both next gen consoles as described in the leaks. Therefore, I don't feel that there is a real reason for them to have huge amounts of memory installed.

If I look back up at the table I compiled and think about the timelines for availability combined with the expectations of what each console is going to be doing (probably going to be a separate post!) I don't think either console will have a clamshell design for the memory due to the issues I pointed out above. Additionally, I think 3 GB GDDR7 modules are the most likely to be used and available in sufficient quantity to satisfy mass production of both consoles.

For the Xbox, I think 18 GB would be enough to satisfy the demands of next gen games (maybe something like 14 GB for games and 4 GB for OS - even if it's going to be running PC applications as part of its remit, this quantity shouldn't pose a problem.

The PS6 wouldn't make sense regressing to 15 GB total capacity. However, if we look at the PS4 Pro and PS5 Pro, they have additional, slower memory that seems to share part of the memory bus to the main portion of memory. While it hasn't been leaked, I could see a similar situation play out for this next gen console instead of waiting for the mid-gen refresh. A console with 15 GB for games and 4 GB separate memory for the OS would work really well.

Yes, I could be wrong about all of this - I've been wrong before when it came to speculation. I am just struggling to see why Microsoft or Sony would vastly inflate the memory capacity - especially if developers are unlikely to make use of it. After all, current gen consoles will still exist, in parallel with PC gamers, too. Are we really going to be expecting 20 - 30 GB games when 12 - 16 GB VRAM GPUs are the limit on PC?

Additionally to this, why would they handicap the performance of their systems by running heavy AI workloads on the system? This is the Kinect situation all over again! Didn't we learn anything from that?

We've never seen that sort of a move before in the gaming space... I'd like to see someone give a good reason why we should expect a change, now.
at

No comments:

Post a Comment