[MLiD] PS6 Dockable Handheld Leak: AMD Canis Specs CRUSH XBOX Ally X!

Same for nvidia. In fact they are first.

First desktop GPU with 3GB modules launch Q4 2025. You really expect 2x bigger modules in 2027? I expect 4GB to be latest and greatest in 2027.



Yeah, they won't split RAM like that.
So you want them to fuck up next gen with 24GB of RAM?

Then again, we are already going to be held back the handheld anyway, so I guess 24GB would work.
 
Last edited:
So you want them to fuck up next gen with 24GB of RAM?

We are already going to be held back the handheld anyway, so I guess 24GB would work.

32GB is what those consoles should have, but I fully expect 24GB.

Remember that PS6 is rumored to be cheaper than PS5 Pro. But that rumor was before it increased in price, lol.
 
I don't expect more than 18GB myself 32 seems a pipe dream for a dedicated console, especially with a dedicated APU sharing that load.
You guys are ignoring parts of the leak.
wqSng0rd5YJNBcqB.jpg


If the PS6 is A.I. heavy (not the AI upscaling portion), it needs more RAM than normal.
 
40GB for PS6 and 32GB for a 1080p handheld is a laughable assumption, theres 0 chance they will come with that nor will they need that much. Very few games (if any at all) need more then 16gb at native 4k. PS6 wont even be a native 4k console and will be pushing the latest FSR heavily.

Regarding LLMs, i think this will be a next next gen thing, no need to be pushing AI this heavily, nvidea will probably lead the charge in the coming years but it will be super jank and uncanny and wont reach full stride till the end of the PS6 generation, AMD can play catch up when the time comes.

PS6 im betting with have 24gb with 18-20 reserved for games. (more than enough imo)
 
Last edited:
You guys are ignoring parts of the leak.
wqSng0rd5YJNBcqB.jpg


If the PS6 is A.I. heavy (not the AI upscaling portion), it needs more RAM than normal.
Maybe you're right but in the context of video games I equate AI to frame generation/upscaling/and image processing not logic and workflow. I could be wrong but the problem is going to be costs. Adding all this ram and AI is costly and they can't make 1000 consoles and expect to have millions of units sold a year.

Also RAM has always been an issue with developers and consoles, there is never enough
 
Last edited:
40GB for PS6 and 32GB for a 1080p handheld is a laughable assumption, theres 0 chance they will come with that nor will they need that much. Very few games (if any at all) need more then 16gb at native 4k. PS6 wont even be a native 4k console and will be pushing the latest FSR heavily.

Regarding LLMs, i think this will be a next next gen thing, no need to be pushing AI this heavily, nvidea will probably lead the charge in the coming years but it will be super jank and uncanny and wont reach full stride till the end of the PS6 generation, AMD can play catch up when the time comes.

PS6 im betting with have 24gb with 18-20 reserved for games. (more than enough imo)
It's not like Sony hasn't been exploring and pushing A.I in games beyond just upscaling.

Gran Turismo Sophy is a good example.

I suspect Sony would push this even further with the PS6 generation.

But what many fail to understand, next gen isn't just about PS5 fidelity at higher resolutions and frame rate. Devs would go for higher fidelity regardless.

Even without LLMs, 24GB isn't enough. The PS5 Pro for example had to increase to 18GB.
 
Maybe you're right but in the context of video games I equate AI to frame generation/upscaling/and image processing not logic and workflow. I could be wrong but the problem is going to be costs. Adding all this ram and AI is costly and they can't make 1000 consoles and expect to have millions of units sold a year.

Also RAM has always been an issue with developers and consoles, there is never enough
Did you not see MLiD cost breakdown?
 
You guys are ignoring parts of the leak.
I call it bullshit as there is no point to run LLM locally on console. And practical tasks that gaming console could run require like 1-3% of that space. DLSS4 transformer model (i.e. same basis model as LLM but tailored for specific task) is 300Mb not 3-10Gb.

If the PS6 is A.I. heavy (not the AI upscaling portion), it needs more RAM than normal.
It's a HUGE waste of resources and could and should be put into cloud portion.
Specialized LLM (transformers) are small and there are no point to put general purpose "jack of all trades, master of none" LLM onto console. You don't need game to chat about amazonia butterflies, it's completely stupid waste of resources.

edit - btw to run single query on 10GB LLM will take at least 5-10 sec full load, making it not applicable to ~any~ gaming purposes
 
Last edited:
It's not like Sony hasn't been exploring and pushing A.I in games beyond just upscaling.

Gran Turismo Sophy is a good example.

I suspect Sony would push this even further with the PS6 generation.

But what many fail to understand, next gen isn't just about PS5 fidelity at higher resolutions and frame rate. Devs would go for higher fidelity regardless.

Even without LLMs, 24GB isn't enough. The PS5 Pro for example had to increase to 18GB.
Horrible example of the GT Sophy application, its a model that has already been trained and isn't some massive LLM running locally with large vram requirements. Additionally, the same Sophy AI runs fine on the base PS5.

If i'm thinking about game design and how LLMs will be integrated in future use cases (fe NPC conversation). I see much smaller more specific models that can be loaded in out of a limited pool of vram at will depending on the context/scene. Having large vram/unified ram pool in a consumer grade device meant for mass consumption is not feasible from a BOM stand point.

Predicting:
PS6 Portable 16-24 GB unified
PS6 24-32 GB unified
 
Last edited:
That will be Medusa Premium with 24 CU RDNA5, though that will have the same issue of current Z1E/Z2E handhelds of just using an off-the-shelf laptop APU which means running at 28W (or higher) to get optimal performance. Canis is proper handheld APU that runs optimally at 15W.

Dang bro....you got me kinda hyped for this joint now :messenger_tears_of_joy:
 
WTF is 3-10 Gb AI requirement. It sounds like invented bullshit. Why console would need to run full scale general-purpose, shit-quality LLM models and not just tailored models like DLSS4, which takes ~300Mb at 4K

EXACTLY!!!! WTF is going on? No way Sony is this stupid.
 
Horrible example of the GT Sophy application, its a model that has already been trained and isn't some massive LLM running locally with large vram requirements. Additionally, the same Sophy AI runs fine on the base PS5.

If i'm thinking about game design and how LLMs will be integrated in future use cases (fe NPC conversation). I see much smaller more specific models that can be loaded in out of a limited pool of vram at will depending on the context/scene. Having large vram/unified ram pool in a consumer grade device meant for mass consumption is not feasible from a BOM stand point.

Predicting:
PS6 Portable 16-24 GB unified
PS6 24-32 GB unified
I used GT Sophy to say Sony was very interesting in AI in games back then, so now they'll go further than that and ues LLMs on the PS6 generation.

So why are you still predicting upto 32GB?
 
Last edited:
I call it bullshit as there is no point to run LLM locally on console. And practical tasks that gaming console could run require like 1-3% of that space. DLSS4 transformer model (i.e. same basis model as LLM but tailored for specific task) is 300Mb not 3-10Gb.


It's a HUGE waste of resources and could and should be put into cloud portion.
Specialized LLM (transformers) are small and there are no point to put general purpose "jack of all trades, master of none" LLM onto console. You don't need game to chat about amazonia butterflies, it's completely stupid waste of resources.

edit - btw to run single query on 10GB LLM will take at least 5-10 sec full load, making it not applicable to ~any~ gaming purposes
Putting it in the cloud would introduce latency issues.

PS6 AI & LLM Memory Slice with 36 GB System Memory should look like this.
  • LLM Weights in VRAM/RAM → 4–5 GB (for on-device models, ~7B–13B parameters quantized)
  • Context Buffers / Token Memory → 1–1.1 GB (for ongoing conversations, story states, NPC-specific context for dialogue & behavior, open-world state, procedural world generation)
  • Inference Buffers → 900 MB (temporary compute space for NPC responses, story decisions, procedural generation computations)
Subtotal ≈ 5.9–7 GB

For me, next gen should be more then just running PS5 games at a high resolution and frame rate.

LLMs enhance NPC intelligence: dynamic dialogue, emergent behaviors, contextual awareness, and story generation.

Memory & compute implications: NPC-driven LLMs consume context buffers, inference buffers, and weight memory, feeding the AI pipelines for real-time behavior.

This makes AI-heavy games feel alive and reactive, rather than static or scripted. But i get it if you would to continue playing static and scripted games just because you want the console to be $50 - $100 cheaper.

I swear I'm the only one that wants a proper next gen leap on this forum.
 
With ps5 sales dropping a little aggressively half way ish into the gen, I wonder if they are dropping in 2026 instead or early 2027.
Can they easily just force assembly line for "cutting edge tech" to happened one year earlier just because they want to? Sounds like a recipe for failure rate disasters.
 
The same delusions from the Windows Phone era, that Microsoft's magical and mysterious software would make everything run on Windows Phone, PC, and Xbox without needing a port.

Account created in 2018 that only became active in 2025.

Don't you Xbox users ever get tired? Don't you have anything to do with your life?
You really like this Windows Phone comparison (I've seen you bring it up in every conversation surrounding Xbox 's next platform). I don't think it's as good of analogy as you do, but of course, you are allowed to have opinions.
SQX6v77Pv3wo62yH.jpg
 
Putting it in the cloud would introduce latency issues.
As if running it locally would not introduce them. The amount of computations required for 10b LLM put it way out of console hardware tier.

For me, next gen should be more then just running PS5 games at a high resolution and frame rate.
It should be first and foremost resolution, fps, effects and other **practical** stuff
Not some random bullshit no one care about.
Expert systems those can imitate a lot of stuff mentioned are readily available on PS4 (or even PS3) hardware - no one really care about implementing them. Carefully scripted "very human-like behavior" bots were crafted for UT2004 - guess what - no one wants them. People like to play with real people and for bots they want them to be bots - simple and predictable.

LLMs enhance NPC intelligence: dynamic dialogue, emergent behaviors, contextual awareness, and story generation.
Nobody care about this shit.
Nobody care even for non-linear stuff, BG3 has the same "false choice" as PlaneScape : Torment had (two games hailed for "choice") where you literally choose your ending 5 mins prior to the end.
People don't really like a random stuff that more often than not goes against them. It make them frustrated and unhappy, and games are leisure aka pleasant activities for mass market.

Memory & compute implications: NPC-driven LLMs consume context buffers, inference buffers, and weight memory, feeding the AI pipelines for real-time behavior.
This makes AI-heavy games feel alive and reactive, rather than static or scripted. But i get it if you would to continue playing static and scripted games just because you want the console to be $50 - $100 cheaper.
I swear I'm the only one that wants a proper next gen leap on this forum.
Even for this shit you don't fucking need LLM - you need transformer trained for specific task in specific environment. WTF your LLM should be able to talk about Amazonia butterflies, Fourier transform or RNA replication when it's about dark fantasy setting of world dying - it doesn't need 99% of those topics and weights related. Get it distilled and trimmed and it'll be DLSS4 size, convenient for specific task and knowledge
 
Last edited:
You really like this Windows Phone comparison (I've seen you bring it up in every conversation surrounding Xbox 's next platform). I don't think it's as good of analogy as you do, but of course, you are allowed to have opinions.
SQX6v77Pv3wo62yH.jpg

The rhetoric is the same. Well, I believe that Microsoft in all areas has the same fan profile... Always with megalomaniacal fantasies of world domination.
 
If it will be cheap and with auto downscaling for the portable version a game instead of obligating developers to add another work load to create a lower spec version of a game( just like what happened to Series S, which affected game development and optimization ), then I might buy it.

But Sony PS must not forget the most important one, and that this exclusive games with true exclusivity. This what made Nintendo Switch 1 and now Switch 2 so successful . And bring back Sony Japan studio AA and A exclusive games.
 
Does this mean PS6 is mid or is Sony going to have two kinda-close spec targets for the next 10 years? Because I'd bet a lot of third-party PS6 games could run on either PSP3 or PS6, but you don't want to limit that first party development to having to make their games run on the PSP3.
 
Does this mean PS6 is mid or is Sony going to have two kinda-close spec targets for the next 10 years? Because I'd bet a lot of third-party PS6 games could run on either PSP3 or PS6, but you don't want to limit that first party development to having to make their games run on the PSP3.
Possibly a scaling technology without obligating developers two create two different version or downgrade a specific game which is one of the main culprit for bad optimization this generation especially on multiplatform games(in the case of xbox series S) and developers not maximizing the full potential of a specific hardware.
 
Sony always following Nintendo, no shit... The good thing is that probably many people will eat crow since it won't be much more powerful than Switch 2... Hardware being this powerful at most for a while is good imo, it gives devs time to polish their tools and start over.
The nintendo brainrot is so strong. Open a window and get some fresh air, your brain has gone stale.
 
What if its 12-16GB of cheap slower ddr5 and 16-20GB of faster GDDR7 simply for cost reduction?
I really hope not. That will just complicate development, potentially will make shrinking the system tougher, and make the whole thing more clunky.

I could see having like 4GB DDR5 for OS or whatnot, but more would be pointless and counterproductive. IMO, of course.
 


Pictures, specs, and the estimated cost of the powerful APU inside PlayStation 6 "S"!

Chapters:
0:00 Pictures of AMD Canis based on Dimensions & Specs!
1:02 PS6 Handheld Specs Overview
2:51 BEWARE of FAKE Rumors
4:37 Sony MASSIVELY upgraded the Specs!
5:29 PS6 Canis & PS6 Orion -Cost Estimate
12:30 This wasn't Easy! Subscribe & Join Patreon!!!
12:30 Game Dev Testimonials on Required RAM
17:57 PS6S Performance & Full Specs Leak
25:30 Sony wants PS4 Gamers to UPGRADE

c7ixwZqtoIbHvOVv.jpeg


PA3p6Nh3ODGO0CA5.jpeg


PS6 "Consoles" estimated cost:-

G7gavrf0kANmzQ5F.jpeg


Anonymous Game Dev Testimonials on Next-Gen RAM Requirements:-

6xkTS88phNS0i8zF.jpeg




Comedy Central Lol GIF by Awkwafina is Nora from Queens

Homie it's an RDNA5 handheld while the Ally is on RDNA3. That's like saying the PS5 crushes the Pascal Titan.. of course it does the PS5 is 2 generations of technology ahead of it.
 
Nothing has enough performance to run next-gen games and an LLM at the same time so this argument is irrelevant.
I'm just going by what he's leaking.
I mean he leaked both the specs of Magnus and Canis, and die diagrams on top of that.

I don't believe he would talk about LLMs unless he saw it in the documents or notes he has.
 
Last edited:
Nothing has enough performance to run next-gen games and an LLM at the same time so this argument is irrelevant.
I brought up this patent a few weeks ago as an example of ML potentially being used to do more than just upscaling on hardware.


The way I see it first there were pixels then vectors and now AI vectors (GAN texture model). Pixels, vectors, latents — each step a higher-order abstraction, pointing toward generative scene reconstruction. I mean the way we are headed you could see AI reconstructing entire scenes based on a design language matched to a trained model, but to begin with you have texture reconstruction as covered in the patent. LLM's are just a single possible use case of on board ML capability. Anyway don't get hung up on this coming generation having a LLM focus. I suspect thats either going to be cloud based for now and/or the following generation on board.
 
Last edited:
Possibly a scaling technology without obligating developers two create two different version or downgrade a specific game which is one of the main culprit for bad optimization this generation especially on multiplatform games(in the case of xbox series S) and developers not maximizing the full potential of a specific hardware.
Secret sauce strikes again
Tools (tm) might help but there will be no automatic scaling down technology (and devs prefer to scale down and not scale up), so testing profiling and adjustment still be required.

The way I see it first there were pixels then vectors and now AI vectors (GAN texture model). Pixels, vectors, latents — each step a higher-order abstraction, pointing toward generative scene reconstruction. I mean the way we are headed you could see AI reconstructing entire scenes based on a design language matched to a trained model, but to begin with you have texture reconstruction as covered in the patent. LLM's are just a single possible use case of on board ML capability. Anyway don't get hung up on this coming generation having a LLM focus. I suspect thats either going to be cloud based for now and/or the following generation on board.
LLMs are big because they are general purpose and tries to store humanity whole data. There is no need for that for tailored AI designed for specific tasks that make them much smaller and faster.
Size and speed for ML directly correlated (it's just huge amount of multiply+add operation against stored vectors of weights) and 3Gb network will take x10 times to proceed than 300Mb network.
 
Secret sauce strikes again
Tools (tm) might help but there will be no automatic scaling down technology (and devs prefer to scale down and not scale up), so testing profiling and adjustment still be required.
Yes there is. A wrapper. Same thing they do with backwards compatibility and why you can play a 360 game at 4K. In this case it'll be the opposite: divide the original resolution.

Edit: never mind. I just realised that you're responding to some delusional soul that thinks that the PS6 portable will run PS6 games, not scaled down PS5 games.
 
Last edited:
Top Bottom