PlayStation 4K - Codename Neo - Technical analysis

[3dilettante]

http://www.eurogamer.net/articles/2016-06-10-sony-confirms-playstation-4k
So not a transparent hardware compatibility, enough that some games may not play on Neo.

It will be interesting to see whether Chipworks-now-TechInsights will have a go at the Neo SOC at launch like it did with Orbis.
I'd hope "supporting" Neo means using its features and benefits, rather than just being functional on it. The whole established software base would be as lost as if Neo were a full next-gen if the games couldn't play on it

 

[milk]

It will be interesting to see whether Chipworks-now-TechInsights will have a go at the Neo SOC at launch like it did with Orbis.
I'd hope "supporting" Neo means using its features and benefits, rather than just being functional on it. The whole established software base would be as lost as if Neo were a full next-gen if the games couldn't play on it

I had your interpretation too, and I have a very large palm prepared for an epic forehead slap if its not the case.

 

[3dilettante]

If Neo is using Polaris (built-in assumption of 14/16nm) as the foundation for its GPU, we can take into account that the RX480 draws at most 150W at 1.2GHz versus the rumored 911 MHz for Neo. Neo has a good chance of drawing significantly less power than Orbis.
Possibly, if the decoding block and omissions in system management are also improved, it won't draw the so much power while streaming or sitting on the dashboard.

 

[hesido]

http://www.eurogamer.net/articles/2016-06-10-sony-confirms-playstation-4k
So not a transparent hardware compatibility, enough that some games may not play on Neo.

I do think that he probably means supporting the features. NEO should run all PS4 library with 0 developer intervention. If not, what a stupid thing Sony is doing. Even worse situation than most of us here thought.

 

[Shifty Geezer]

It's a but fuzzy but you may be right. There was mention in an article about when there's a bug in the Neo code path, devs wouldn't allow users to drop to the PS4 code path and would have to fix the bug. That kinda supports your view, but then the claim that games can be made to work with little effort and some games might not run means we aren't going to be putting old game discs in and be playing (in higher res!).

 

[DieH@rd]

That kinda supports your view, but then the claim that games can be made to work with little effort and some games might not run means we aren't going to be putting old game discs in and be playing (in higher res!).

IMO, I expect that 99% [if not 100%] of PS4 games will work fine on Neo, Neopatched or not. If some crazy chinese indie game does not work [or delisted P.T. ], I can live with that.

 

[Allandor]

IMO, I expect that 99% [if not 100%] of PS4 games will work fine on Neo, Neopatched or not. If some crazy chinese indie game does not work [or delisted P.T. ], I can live with that.

Well, many PS4 games are writter lower to the metal than e.g. xbox games. With this in mind, they are somehow limited in options. A new GPU architecture could change some things. Things that were fixed functions before (and were almost free) could be costly with a new architecture (happend a few times in the past).
So if Sony is lucky the GCN 2.0 architecture is 1:1 backwardscompatible, if not, they might have a problem and need patches for old games. It's the same for the CPU. It is still x86 for sure, but if the Caches have changed or anything, it could cost some framerates. But that could be "fixed" by a higher clock-rate.

 

[DieH@rd]

Devs have already said that PS4 API is low level, but it is still abstracted. They are not writing completely to the metal, almost same as with DX12/Vulkan [who still have to account for few different vendors].

I would be SHOCKED if Sony did not know about AMDs capability to produce whatever hardware update they wanted down the line. If they have ounce of brain, they would have created their API/Drivers/OS to be forward compatible with future X86/GCN updates [and to allow stackable API updates that can expand capabilities when new hardware is introduced]. They should have all this planned out 2 years before PS4 was released, when they started working on their OS/Drivers/API.

Heck, the first thing AMD shoud have said at the meeting when they offered their services to Cerny was: "We can offer stable and upgradeable X86 foward-compatibile platform that can last a century."

 

[milk]

Devs have already said that PS4 API is low level, but it is still abstracted. They are not writing completely to the metal, almost same as with DX12/Vulkan [who still have to account for few different vendors].

I would be SHOCKED if Sony did not know about AMDs capability to produce whatever hardware update they wanted down the line. If they have ounce of brain, they would have created their API/Drivers/OS to be forward compatible with future X86/GCN updates [and to allow stackable API updates that can expand capabilities when new hardware is introduced]. They should have all this planned out 2 years before PS4 was released, when they started working on their OS/Drivers/API.

Heck, the first thing AMD shoud have said at the meeting when they offered their services to Cerny was: "We can offer stable and upgradeable X86 foward-compatibile platform that can last a century."

I hope to god you are right

 

[Pixel]

The GPU is weakened by memory bandwidth contention. Unless they make some significant revision of the jaguar cpu architecture by upping the L2 cache from 2*4MB to something like 2*8MB, which almost certainly wont happen.

at 10GB/s cpu bandwidth utilization you roughly get 105GB/s * 124% = 130GB/s bandwidth available to the NEO gpu on a 36 core (CU) polaris gpu. Couple that with lossless delta color compression introduced with GCN1.2 Tonga (R9 285) its still far behind the RX 480.

Compare that with RX480 with ~200GB/s real world bandwidth to the GPU (~80% 256GB/s theoretical peak), and NEO only has 65% of the 480 real world memory bandwidth.

Even with 200GB/s memory bandwidth may be constricting the RX480. As AMD has shown diagrams that GPU performance improvements will be slowed by GDDR5.

 

[Grall]

@Pixel Console hardware always has compromises built into it. That CPUs chew up memory bandwidth is unavoidable in an UMA system anyhow, it won't matter how big you make the caches (unless they more or less match main RAM, and that would be kind of stupid wouldn't you say ); the CPU is always going to need some bandwidth, and that will by necessity have to come out of the piece of the pie the GPU would otherwise been able to eat up.

...Depending on workload, of course.

 

[Pixel]

If games like Batman Arkham Knight & Witcher 3 & Fallout4 & Future GTA6 & Future RDR2 sequel etc etc pushed ps4 gpu memory bandwidth to 90GB/s or more (almost a certainty), doubling the fps on Neo won't be possible with only a 24% increase in memory bandwidth. You can tell by the anisotropic filtering limitations and texture size limitations this generation that they are pushing against ps4's ( & also Xbox One's) memory bandwidth limitation in many situations.


Compared to the 320 GB/s theoretical peak for Scorpio. 46% higher memory bandwidth than Neo should provide not only the shaders but also the memory bandwidth to double fps on many games compared to the regular consoles. Neo is setting itself up to be in an odd place where it can't double the fps due to cpu and memory bandwidth limitations, but can only offer cosmetic improvements to games.

 

[AlNom]

Delta compression might help somewhat assuming modern GCN arch.

 

[BRiT]

Delta compression might help somewhat assuming modern GCN arch.

But both consoles will have that, so not like it can be used to make up the gap.

 

[kalelovil]

Compared to the 320 GB/s theoretical peak for Scorpio. 46% higher memory bandwidth than Neo should provide not only the shaders but also the memory bandwidth to double fps on many games compared to the regular consoles. Neo is setting itself up to be in an odd place where it can't double the fps due to cpu and memory bandwidth limitations, but can only offer cosmetic improvements to games.

No matter its power, Project Scorpio is also only going to offer 'cosmetic improvements' if Microsoft sticks to their promise of forwards compatibility from the Xbox One.

(Or as its the runt of this generation-and-a-half of consoles, demanding games are allowed to run at a crawl on the Xbox One)

 

[Pixel]

No matter its power, Project Scorpio is also only going to offer 'cosmetic improvements' if Microsoft sticks to their promise of forwards compatibility from the Xbox One.

(Or as its the runt of this generation-and-a-half of consoles, demanding games are allowed to run at a crawl on the Xbox One)

They said it themselves "it will offer the best framerates" "60fps". Forwards compatibility definitely includes the allowance of framerate differences particularly in singleplayer portions of games. How is that breaking its promise of forwards compatibility?
The definition of forwards compatibility means the new software that supports the newer system will still run on older hardware.

Some developers may offer lobby options to chose between segregated multiplayer games at 60fps with other Scorpios, and unsegregated which maintains 30fps (or identical fps) with the much much larger regular XO playerbase.

 

[sebbbi]

So if Sony is lucky the GCN 2.0 architecture is 1:1 backwardscompatible, if not, they might have a problem and need patches for old games. It's the same for the CPU. It is still x86 for sure, but if the Caches have changed or anything, it could cost some framerates. But that could be "fixed" by a higher clock-rate.

Polaris is GCN 4.0. AMD already did some changes to GCN 3.0 (Tonga/Fiji) that broke direct backwards compatibility to GCN 1.0 and 2.0. Read this document: http://amd-dev.wpengine.netdna-cdn..../07/AMD_GCN3_Instruction_Set_Architecture.pdf. It has a list of deprecated shader instructions among other breaking changes. So you'd at least need to patch the shader microcode at runtime if it contains these instructions. Polaris hopefully also has front end changes to catch up with Nvidia (at least there are lots of rumors about this). It might not be trivial to patch everything at runtime. But we have seen (last gen) PPC code translated automatically to x86 in the past with pretty good results (and last gen VLIW shader code translated to scalar architectures), so this shouldn't be a roadblock by any means.

CPU obviously is not a problem, as long as it has at least 8 cores. x86/64 is a well standardized architecture. Zen has much faster and bigger caches than Jaguar, so I wouldn't be worried about cache performance.

 

[RDGoodla]

If games like Batman Arkham Knight & Witcher 3 & Fallout4 & Future GTA6 & Future RDR2 sequel etc etc pushed total ps4 system memory bandwiidth to 90GB/s or more (almost a certainty), doubling the fps on Neo won't be possible with only a 24% increase in memory bandwidth. You can tell by the anisotropic filtering limitations and texture size limitations this generation that they are pushing against ps4's ( & also Xbox One's) memory bandwidth limitation in many situations.


Compared to the 320 GB/s theoretical peak for Scorpio. 46% higher memory bandwidth than Neo should provide not only the shaders but also the memory bandwidth to double fps on many games compared to the regular consoles. Neo is setting itself up to be in an odd place where it can't double the fps due to cpu and memory bandwidth limitations, but can only offer cosmetic improvements to games.

How about GTX970? It can run the above games @1080p max setting & close to 60fps, while it doens't have
huge bandwidth.

Of course I agree that the best way for Neo is to choose a new CPU.

 

[Pixel]

How about GTX970? It can run the above games @1080p max setting & close to 60fps, while it doens't have
huge bandwidth.

Of course I agree that the best way for Neo is to choose a new CPU.

In the PS4 at 5GB/s cpu usage the gpu has ~120GB/s memory bandwidth available to itself.
In games that push cpu memory bandwidth like the DDR4-overclocking-loving-Fallout4 games that utilized 10GB/s system memory bandwidth, the gpu only has access to ~ 105GB/s.

Remember in a unified memory system the higher the fps the more memory bandwidth the cpu requires and the more bandwidth the cpu requires the more contention it creates. Regardless of how much system memory the cpu takes in a split memory pool system, it will have little impact on the dedicated video memory bandwidth available to a discreet gpu.

 

[Arwin]

How about GTX970? It can run the above games @1080p max setting & close to 60fps, while it doens't have
huge bandwidth.

Of course I agree that the best way for Neo is to choose a new CPU.

Actually it does have a pretty decent bandwidth that it also doesn't need to share with CPU? Point is that 4K requires a lot more, or a pretty serious downgrade of visual effects (I tried a bunch of games on 3 screens)


Sent from my iPad using Tapatalk