Predict: The Next Generation Console Tech

Status
Not open for further replies.
The ACEs are effectively the workload setup and dispatch logic that is present in the graphics pipeline, minus the graphics-specific portions. It's somewhat ambiguous as to whether one of the ACEs in Tahiti is the graphics pipeline, but the graphics pipeline can run compute and graphics commands, while the ACEs are limited to the compute subset.

General cache coherency is more the province of the cache subsystem, which GCN has linked most parts of the chip to, including the parts of the graphics portion not explicitely kept separate.

The more interesting idea would be if there were CPU cores were under the ACEs as well as CUs, which at least in theory could be a future direction to take things. ACEs assign contexts and set up the register allocation and flag setup per wavefront assignment, which at a high level is what loading a CPU thread is, absent hardware accelleration of the process.

edit: One could also make an x86 core take the place of an ACE, for a more software-driven approach.

What is the usual performance hit between coherent and incoherent memory access. Some of the rumors I have seen cite 2 different bandwidth numbers. One for "coherent bandwidth" (lower), and one for incoherent. Assuming 8 Jagcores and 1 laptop GPU, what does incoherent access mean ? (Between the Jag cores or between the CPU and GPU ?).

thats not true,

if you read posts of beginning 2012 on this very thread, I was one of the few expecting 4 Gb of GDDR5 RAM for PS4, and I was criticized as being unrealistic...(too expensive)

I was also one of the very few to expect a top of the line laptop GPU for ps4, the idea was attacked furiously with the argument that "laptop top of the line GPUs cant be used for consoles because they cost too much and they come on limited quantity"

There there... [pass hanky to doctor]


The ps3 if DF got it right looks like an interesting device.

With regard to the "secret compute sauce" I could bet on it in fact being part of the main GPU.
We know that for instance the HD 7970 has 2 ACE (asynchronous compute element) so Pitcairn.

Might be good to give a link to a solid ACE resource/reference on the net.
 
What is the usual performance hit between coherent and incoherent memory access. Some of the rumors I have seen cite 2 different bandwidth numbers. One for "coherent bandwidth" (lower), and one for incoherent. Assuming 8 Jagcores and 1 laptop GPU, what does incoherent access mean ? (Between the Jag cores or between the CPU and GPU ?).
That sounds like a reiteration or an evolution of how current AMD APUs are structured. The GPU memory controller is kludged onto the CPU northbridge IMC, and is linked with a set of channels. It's one of the less than ideal facets of current off-the-shelf designs. It could be different now for a custom chip, though.

The larger is the non-coherent one, which can afford to do more because it doesn't need to query the cache status of the lines it is reading, allowing it to use the full bandwidth of the memory controller.
The GPU could very readily swamp the CPU memory hierarchy if it wanted to, so the coherent bus is smaller.
Depending on the core interconnect implementation, the same may not always be said of each CPU, and in general CPUs don't burn bandwidth like GPUs--in part because they shouldn't, shouldn't need to, and probably cannot physically have that many accesses in flight.
 
Maybe you will be able to use skype while gaming (even though you have your PC right beside you). That sounds like a great way to spend resources.

*I doubt there is 3GB of RAM dedicated for OS. That seems horrible even for Nintendo standards

With that much RAM reserved, you'd have to assume some CPU cores are dedicated to OS tasks.

Possible reasons:
-Use skype for all video and voice chat on the new Xbox
-Facebook, twitter, email + other social media constantly running with notifications that can be checked
-Persistent Live Avatar world that no one cares about
-DVR features running in background
-High-res animated 3D blue-screen-of-death
-Fast switching between games and full dashboard apps (no more quitting your game to change console settings, check what got released on Xbox Video), or some form of multitasking in OS
-Full featured media player running in background
-Huge OS bloat that leads to poor performance and wasted resources
-Kinect processing (Audio, Video, skeletal junk) handled by OS
-Runs Windows 8 RT
-Full-featured Internet Explorer with tabs (with multitasking this could be nice for OCD gamers that use game guides)

Think Windows 8 tablet with fixed hardware reserved for running games. I'm guessing they're chasing more the "home entertainment" crowd, than just pure gaming. Just my 2 cents.
 
thats not true,

if you read posts of beginning 2012 on this very thread, I was one of the few expecting 4 Gb of GDDR5 RAM for PS4, and I was criticized as being unrealistic...(too expensive)

I was also one of the very few to expect a top of the line laptop GPU for ps4, the idea was attacked furiously with the argument that "laptop top of the line GPUs cant be used for consoles because they cost too much and they come on limited quantity"

The only thing where I was hugely mistaken is about CPUs, I expected a derivative of IBM power8 CPU for nextgen consoles, I was wrong, I didnt expect sony or microsoft going for weak CPUs...its the only disappointment for me...

You can't think of desktop and laptop GPUs as separate. The mobile 7970M is a desktop 7870 binned for low power. If a die can't meet the 7970M requirements, it becomes a 7870. If a die can't meet the ps4 requirements, it goes in the garbage. It's also a no margin game whereas there's a lot of margin in mobile cards. The scenarios simply aren't compatible. He used 7970M for convenience sake.

So you're saying the 75W is the whole card? I'm not sure I follow. He accounts for RAM etc later on.

The whole card is 100W. The RAM isn't a whole 25W.
 
Unlikely, but I so very much want to see Last of Us on the new box, plus proper 3D. Can someone develop it in parallel with NaughtyDog ? 8^d
 
I haven't been following much of this off late, titan quest kind of stole my life...

What is the word on the street with regard to Durango? Is it still a whole AMD design?
A join effort effort IBM /AMD / in house?

---------------
Other than that the patents posted by liquidboy (?) were interesting by self, though possibly ot :LOL:
It looked to me like a really general purpose take of the GPU concept.
A generic command processors feeding different types of elements, interesting.
 
Alright, but if the 75W is just the chip, why is it wrong? He lumps the overhead of everything else to get an upper limit estimation anyway.

I don't know what the chip is. I simply know the card is 100W and the RAM isn't 25W. He's going to get into trouble assuming 75W for the chip because that's an aggressively binned part. He may get that back by having 2 less CU's and GCN2 being more efficient, but I would lump the GPU and RAM and assume at least 100W for the pair.


Unlikely, but I so very much want to see Last of Us on the new box, plus proper 3D. Can someone develop it in parallel with NaughtyDog ? 8^d

I'm sure they'll throw in some back compat and upscale to 1080p and throw in some x8 AA just for you ;)
 
An xbox next with lots of on board memory(L1/L2 caches, edram, system ram etc) and very large GB/s buses between components will facilitate a RayTracing engine, this is what im looking for in rumoured specs for xbox next .. Time will tell ...

I don't see the next xbox having anywhere near the amount of memory or bandwidth to render an entire scene with a lot of complexity. We don't even have enough memory in our current PCs. A typical rendered scene for RT in film is easily 12-16G. And that is per frame. We are talking about throwing in tessellation for displacement mapping, and very complex shader networks (which should be evaluated every time a ray hits a surface instead of cached).

I just don't see it happening in a console first. I think PCs will get it first and we are still not there yet IMO. I think it would require a new way of rendering from the tried and true paradigm used now. But I could be wrong.. we'll see.
 
Would it be crazy to have 6 SPE for BC, and repurpose them as the DSP, audio, codecs, compression, encryption, etc.. with PS4 games?
(I know it's been talked about a few months back but I can't find it)
 
I don't see the next xbox having anywhere near the amount of memory or bandwidth to render an entire scene with a lot of complexity. We don't even have enough memory in our current PCs. A typical rendered scene for RT in film is easily 12-16G. And that is per frame. We are talking about throwing in tessellation for displacement mapping, and very complex shader networks (which should be evaluated every time a ray hits a surface instead of cached).

I just don't see it happening in a console first. I think PCs will get it first and we are still not there yet IMO. I think it would require a new way of rendering from the tried and true paradigm used now. But I could be wrong.. we'll see.

I suspect we may see it next gen, if there is a next gen.

Would it be crazy to have 6 SPE for BC, and repurpose them as the DSP, audio, codecs, compression, encryption, etc.. with PS4 games?
(I know it's been talked about a few months back but I can't find it)

We were just talking about that a few pages back. Rough calculations tell me a 28nm Cell would be around 60-65 mm^2. Still fairly big.
 
3- Nextgen consoles (especially ps4) first and second gen exclusive games would destroy anything on PC, the historical trend of consoles games being ahead graphically than PC games would continue afterall as I expected too !

so lets see now in fall 2013 - fall 2014, how PC titles could compete with console titles designed with an almost 2 teraflops GPU in mind....(I hope Shifty Geezer is reading this :mrgreen:) when sony reveals this E3 its demo of the next naughty dog ps4 game or guerilla or santa monica or polyphony digital game, jaws will surely drop...

You've already put your theories across in the next gen thread, there's no need to pollute this thread with them as well. If these rumours are true, then PS4 will be sporting around 1/3rd the peak GPU performance of the highest end AMD GPU's at it's launch time. When the 360 launched it was roughly equivilent to the best AMD GPU's (it actually had around 1/3 more FLOPS). So to take these specifications as some kind of proof of a continuation of what happened at the start of last generation (assuming we even accept your dubious interpretation of what that was) is ludicrous at best.

On the subject of the thread, the most intersting thing about these specs for me is the inclusion of what seems like 3 seperate processing blocks all on the same APU. GPU, 8 core CPU and what sounds like a general purpose SIMD unit makes for quite a potent APU! I guess the confirmation of GDDR5 means we're talking about a standard 256bit memory interface which lends further credence to the 192GB/s rumours. If true that's certainly a healthy chunk of bandwidth for what is a fairly modest CPU/GPU combination.

The extra unit dedicated to GPGPU sounds very interesting too. If it is adding another 200GFLOPS to the CPU which I understand itself is pushing around 100GFLOPS then that's a lot of SIMD power. Could be useful for emulating Cell?
 
You've already put your theories across in the next gen thread, there's no need to pollute this thread with them as well. If these rumours are true, then PS4 will be sporting around 1/3rd the peak GPU performance of the highest end AMD GPU's at it's launch time. When the 360 launched it was roughly equivilent to the best AMD GPU's (it actually had around 1/3 more FLOPS). So to take these specifications as some kind of proof of a continuation of what happened at the start of last generation (assuming we even accept your dubious interpretation of what that was) is ludicrous at best.

On the subject of the thread, the most intersting thing about these specs for me is the inclusion of what seems like 3 seperate processing blocks all on the same APU. GPU, 8 core CPU and what sounds like a general purpose SIMD unit makes for quite a potent APU! I guess the confirmation of GDDR5 means we're talking about a standard 256bit memory interface which lends further credence to the 192GB/s rumours. If true that's certainly a healthy chunk of bandwidth for what is a fairly modest CPU/GPU combination.

The extra unit dedicated to GPGPU sounds very interesting too. If it is adding another 200GFLOPS to the CPU which I understand itself is pushing around 100GFLOPS then that's a lot of SIMD power. Could be useful for emulating Cell?

200 GFLOPS CU. :oops:
 
We were just talking about that a few pages back. Rough calculations tell me a 28nm Cell would be around 60-65 mm^2. Still fairly big.
I mean just adding 6 SPE, not the whole Cell. i.e. having the PPE emulated by one of the x86 cores, The "reserved" SPE functionality emulated on another x86 core, and the new GPU emulating the RSX. 6 SPE would be about 42% of the full cell die?
 
Status
Not open for further replies.
Back
Top