PS4 Pro Official Specifications (Codename NEO)

[3dilettante]

The most interesting is GPU. It is not Polaris-based. In fact it is GCN 1.1 with some functions of Polaris.

This answers one of the questions I starting having after the initial leaks for Neo.
What functions it might inherit from Polaris are unclear. The Neo slides did mention new "instructions" although instructions for what is unknown.
Scalar memory operations were significantly changed and might represent an area that could not back-port.
Some of the wavefront and FP16 instructions would be nice to have, but whether there's a place they could fit in the ISA or how much hardware would need to change could have impeded that.

Delta compression would be a significant win, but the full use case would require hardware changes.

Various power-saving measures might be left to Polaris, which could explain some of the power difference. Some of the low-level circuit optimizations would be agnostic to the GPU version, so some could still carry over.
The process difference could be another item that breaks the applicability of Polaris as a reference.

There seems to be an impression that GF's process is less optimal. On the other hand, yet another process jump might mean that Neo isn't as optimal as it could be.

 

[Scott_Arm]

https://twitter.com/PlayStation/status/773600156356866048

"PS4 Pro to take PS4 experience to new heights. GPU based on elements of AMD's Polaris tech, and some beyond #PS4Pro"

Well, reading that again, I guess there's room for it to not be a Polaris GPU, but just inherit some but not all of its features.

Worded cleverly to suggest it's Polaris+, but can also mean Polaris-.

 

[MrFox]

Worded cleverly to suggest it's Polaris+, but can also mean Polaris-.

It's worded to say it's both? They used a comma before the word "and", and I like when they do that, and Shifty taught me everything.

They said the PS4 GPU was improved with features from Polaris and beyond, along with their own custom additions (which right now we only know of the checkboard hardware helper, and it's assumed they also added whatever they did with the original PS4), not that it was a direct derivative of polaris. Journalists are misquoting.

Supposedly Cerny will have a more dev-focused interview soon to explain the details.

 

[Scott_Arm]

Yah, Mark Cerny's exact words on stage were "... and adopted many new features from the AMD Polaris architecture as well as several even beyond it." Somehow, like many others, I thought they said it was Polaris with some customizations.

 

[Jawed]

The earlier leaked slides put it at 36 CUs, at least in Neo mode rather than the backwards compatibility mode. Leaked clock was 911MHz and the now public 4.2 TFLOP value fits.
Neo mode is described as activating shader engines, presumably when the extra 18 CUs are made active.

Since the PS4 already has 32 ROPs with its 2 shader engines, Neo would need 32 ROPs even when the GPU is half-off.

Going for Neo mode will turn on shader engines, and that in normal circumstances bring up a complement of extra ROPs--unless something was changed from business as usual.

36 CUs is too many for 2 shader engines (they supposedly max out at 16 CUs per shader engine, which is what we see in Tahiti and Fiji). So it seems PS4 Pro has 4 shader engines. But, like Polaris 10, it could have only 8 ROPs per shader engine, which would mean it has only 32 ROPs.

Why did Sony go for 9 CUs per shader engine with 16 ROPs? It's pretty unbalanced in comparison with all other GCN GPUs. But if they did it once, perhaps they'll do it again with PS4 Pro.

A 64 ROP Polaris in PS4 Pro would be pretty potent. It would also keep you warm in winter...

 

[Scott_Arm]

It's worded to say it's both? They used a comma before the word "and", and I like when they do that, and Shifty taught me everything.

They said the PS4 GPU was improved with features from Polaris and beyond, along with their own custom additions (which right now we only know of the checkboard hardware helper, and it's assumed they also added whatever they did with the original PS4), not that it was a direct derivative of polaris. Journalists are misquoting.

Supposedly Cerny will have a more dev-focused interview soon to explain the details.

Reading that twitter statement, I'd take it to mean it's a combination of Polaris plus future architecture. But the way it was revealed on stage much more clearly expressed that it was still based in the previous architecture.

 

[MrFox]

Reading that twitter statement, I'd take it to mean it's a combination of Polaris plus future architecture. But the way it was revealed on stage much more clearly expressed that it was still based in the previous architecture.

Think of it as when you see a film "based on elements from a true story".

 

[Scott_Arm]

Think of it as when you see a film "based on elements from a true story".

I'm not convinced. I think it's worded poorly to make it sound like a gpu based on elements of polaris and even newer, but nothing older.

 

[Arwin]

Whatever it is, I preordered one just in case

 

[3dilettante]

36 CUs is too many for 2 shader engines (they supposedly max out at 16 CUs per shader engine, which is what we see in Tahiti and Fiji). So it seems PS4 Pro has 4 shader engines. But, like Polaris 10, it could have only 8 ROPs per shader engine, which would mean it has only 32 ROPs.

Assuming that Neo has 4 engines, and that the full mode turn involves turning all 4 on, then the backwards-compatible mode will not have a full ROP count since there is a static division of ROPs among engines.
Only having 32 ROPs total would mean not having 32 ROPs in backwards-compatibility mode.

Why did Sony go for 9 CUs per shader engine with 16 ROPs?

It's 9 CUs per shader engine with 32 ROPs for the current PS4, which is why I think not doubling the count would require something different than has been done before.

A 64 ROP Polaris in PS4 Pro would be pretty potent. It would also keep you warm in winter...

Since the PS4 Pro is actively targeting 4K with checkerboard upscaling, it reduces the load by half for 4K (being 4x 1080P). Not doubling ROP count would mean ROP throughput would be 1/4 what would be expected of a PS4 scaled to 4K solution.

 

[Jay]

Reading that twitter statement, I'd take it to mean it's a combination of Polaris plus future architecture. But the way it was revealed on stage much more clearly expressed that it was still based in the previous architecture.

if it's Liverpool but with added tweaks and updates, that would have made ps4 compatible mode a lot easier to deliver.

 

[corysama]

http://pc.watch.impress.co.jp/docs/column/kaigai/1019021.html

http://pc.watch.impress.co.jp/img/pcw/docs/1019/021/Visio-DRAMBandwidth.png

That's a fun diagram

 

[Deleted member 87499]

Great article by Hiroshige Goto.

 

[Ryan Smith]

Assuming that Neo has 4 engines, and that the full mode turn involves turning all 4 on, then the backwards-compatible mode will not have a full ROP count since there is a static division of ROPs among engines.
Only having 32 ROPs total would mean not having 32 ROPs in backwards-compatibility mode.

Traditionally ROPs are closely bound to the memory controllers, not the CUs (the exception being Tahiti and its crossbar). Disabling half the ROPs would, under normal circumstances, disable half of the memory channels as well.

 

[3dilettante]

Traditionally ROPs are closely bound to the memory controllers, not the CUs (the exception being Tahiti and its crossbar). Disabling half the ROPs would, under normal circumstances, disable half of the memory channels as well.

The Xbox One and PS4 GPUs show 16 and 32 ROPs interfacing with a 256-bit subsystem, although counting the ESRAM interface would increase the total number of channels to memory storage. I don't recall exactly, but I thought that capability was attributed to a crossbar of some kind as well.

There's a strong link between the render back ends and the rasterizer front ends they are associated with. The slides for Neo's modes have shader engines and presumably their front ends being turned on or off.
We have some precedent for varying the connectivity between ROPs and memory channels, but is there an equivalent for ROPs and shader engine front ends?

 

[Jawed]

GCN doesn't bind ROPs to MCs. That would require that rasterised fragments are moved around the GPU before being shaded. Something that GCN doesn't do.

 

[rekator]

no H.265 encoder like Polaris

Understood that actually, for compatibly with PS4, Sony don't want a H265 encode for the share fonctions… But strange to not included it for future, and reduce bandwidth requirement for Share fonction. At same quality H265 is 50% of H264 bandwidth. Maybe the spec are from a check list functions and not hardware specs?

 

[Arwin]

I think we know that PS4 Pro will have 1080p recording and streaming, so there is at least some difference.

 

[DieH@rd]

I think we know that PS4 Pro will have 1080p recording and streaming, so there is at least some difference.

- 4K screenshots
- 1080p30 gameDVR
- 1080p60 for RemotePlay, SharePlay and online broadcasting [1080p30 for twitch].

 

[Nesh]

I wonder if its worth it to upgrade to Pro if I dont own a 4K TV. If games will see a substantial increase in framerate or graphic fidelity then it might be worth it but so far it doesnt seem to be the common case. Only the exemption.